We address the problem of encoding room impulse responses (RIRs) into high-order Ambisonics (HOA) representations from arbitrary and potentially insufficient or incomplete microphone array measurements. This task is fundamentally ill-posed for microphone arrays with limited spatial capture capabilities, such as irregular or sparse arrays, as classical linear methods fail to reconstruct high-order spatial detail. We introduce a diffusion-based generative framework that models the statistical properties of HOA RIRs. This enables device-agnostic encoding from arbitrary microphone arrays, potentially unseen during data measurement. Our approach incorporates a posterior sampling procedure that enforces consistency between the estimated signals and the measurements while plausibly reconstructing spatial information that is unobservable from the limited measurements alone. Experiments on simulated data demonstrate that our method outperforms linear and neural baselines, achieving accurate HOA RIR estimation up to 12th order. A listening test with binaural renderings, including both simulated and measured RIRs, further confirms that the proposed method yields higher perceptual similarity to reference Ambisonics RIRs than all baselines. The flexibility and accuracy of the proposed framework opens new possibilities for scalable acoustics simulations.
Primary: Reality Labs Research, Meta
All Institutions: Reality Labs Research, Meta, Aalto University, Acoustics Lab
This paper presents a significant advancement in spatial audio processing by effectively leveraging diffusion models to solve the ill-posed problem of high-order Ambisonics encoding from sparse measurements, achieving state-of-the-art perceptual and objective performance.
The paper proposes a novel application of diffusion models to the ill-posed inverse problem of encoding Room Impulse Responses (RIRs) into High-Order Ambisonics (HOA) from sparse or irregular microphone arrays. The core methodological contribution is the integration of a device-agnostic diffusion prior with a posterior sampling procedure that enforces data consistency via a range-projected likelihood term. The use of a hybrid time-frequency/time-domain architecture (NCSN++ backbone with a U-Net refinement stage for early reflections) is a well-reasoned design choice given the distinct temporal characteristics of RIRs. The introduction of a compressed spectrogram distance metric for the likelihood guidance is a technically sound innovation to emphasize weak high-order components. The formulation separates the device-agnostic prior from the device-specific likelihood, which is a strong theoretical foundation for generalization.
The experimental evaluation is rigorous and comprehensive. The authors utilize two datasets: a large-scale internal FDTD simulation dataset and the public Treble-10 dataset. They compare against three strong baselines: Linear Least-Squares, a Time-Dependent Neural Encoder, and a Conditional Diffusion model specific to the device. The inclusion of both objective metrics (EDC, NPM) and a subjective listening test (MUSHRA-style with binaural rendering) provides robust validation. The results demonstrate clear superiority in both perceptual similarity and objective metrics, particularly in preserving early reflections and late reverberation tails. The ablation studies effectively isolate the contribution of the range-projection constraint and the hybrid architecture.
The paper provides sufficient detail for reproduction, including dataset descriptions, model architectures (NCSN++ based), training hyperparameters (AdamW, learning rate, batch size), and specific implementation details like the compressed spectrogram definition. The use of standard libraries (PyTorch) and well-known architectures aids reproducibility. However, the reliance on an internal FDTD dataset for the primary training results limits independent verification of the scale of the results, though the public Treble-10 results offer some ground truth.
The primary limitation is the distribution gap between simulated training data and real-world measured RIRs, which the authors acknowledge and demonstrate leads to performance drops on measured data (Eigenmike-64). The method assumes a highly accurate Array Transfer Function (ATF), which may not hold in real-world deployments with calibration errors. Additionally, the iterative sampling process is computationally intensive, making it less suitable for real-time applications without significant acceleration techniques. The evaluation is currently limited to a single device configuration (Aria Glasses) for the device-agnostic claim, though the method is theoretically general.
This work has significant implications for spatial audio processing, particularly for Virtual Reality (VR) and Augmented Reality (AR) applications where scalable and accurate acoustic simulation is crucial. By enabling high-quality HOA encoding from sparse arrays, it lowers the barrier for capturing spatial audio with wearable devices. It also facilitates the generation of large-scale training data for spatial audio models. The device-agnostic nature of the approach promotes interoperability across different hardware platforms. This paper presents a significant advancement in spatial audio processing by effectively leveraging diffusion models to solve the ill-posed problem of high-order Ambisonics encoding from sparse measurements, achieving state-of-the-art perceptual and objective performance.
Modern text-to-speech (TTS) is dominated by autoregressive (AR) codec language models, whose left-to-right decoding brings latency that grows with utterance length, error accumulation along the committed prefix, and an artificial generation order imposed on the Residual Vector Quantization (RVQ) token grid. We propose Luna-TTS Family, diffusion-language-model-based TTS systems pretrained on 1 million hours of speech across Chinese, English, Japanese, and Korean. The family is built by progressive adaptation of a pretrained AR text LLM, from causal to bidirectional and finally to block-causal attention, and comprises two variants sharing a single tokenizer, data pipeline, and 0.6B backbone lineage. Luna-TTS is fully non-autoregressive: it generates the entire RVQ token grid in a fixed number of parallel refinement steps, with zero-shot voice cloning and speech editing arising natively as infilling. Luna-TTS Realtime, derived by continual training, is autoregressive over blocks of 32 codec frames (1.28s) while denoising each block in parallel; it supports KV-cached blockwise generation and incremental audio delivery, achieving an end-to-end RTF of 0.0240 and 41.6 ms local first-block latency under the warmed serving protocol. An annealed fine-tuning stage adds explicit control over emotion and non-verbal vocalizations (NVVs), and a reinforcement-learning stage applies GRPO with policy ratios computed over the realized denoising trajectory. On Seed-TTS-Eval, Luna-TTS achieves the best results on all four metrics among compared open-source and commercial systems (0.73 CER / 79.7 SIM on test-zh, 1.49 WER / 76.8 SIM on test-en); on the harder in-the-wild CV3-Eval, it posts the lowest Mandarin and English error rates in our comparison. Against leading commercial systems, it achieves the best results on most objective, model-based, and human-rated metrics for NVV and emotion control.
Primary: VUI Labs Research
All Institutions: VUI Labs Research
Luna-TTS introduces a scalable, block-diffusion-based TTS framework that achieves state-of-the-art quality and latency by adapting pretrained LLMs through progressive architectural changes and RL post-training, effectively resolving the latency-quality trade-off inherent in previous diffusion TTS systems.
The paper proposes a novel architecture for Text-to-Speech (TTS) by adapting Large Language Models (LLMs) into Diffusion Language Models (dLLMs). Specifically, it introduces a "block-causal" attention mechanism that allows for streaming generation while retaining the parallel denoising benefits of diffusion models. The approach involves a progressive adaptation from causal to bidirectional and finally to block-causal attention on a pretrained AR text LLM. A key technical contribution is the application of Group Relative Policy Optimization (GRPO) directly over the realized denoising trajectory, optimizing for content correctness and speaker similarity. The use of a semantic-distilled RVQ tokenizer where the first codebook is anchored to linguistic content is a significant methodological detail that facilitates the diffusion process. The transition from fully parallel masked diffusion to block-autoregressive diffusion for latency reduction is a well-reasoned engineering solution to a known bottleneck in diffusion-based generation.
The evaluation is comprehensive, covering both objective metrics (CER, WER, SIM) and subjective/human-rated metrics for emotion and non-verbal vocalization (NVV) control. The paper claims state-of-the-art performance on Seed-TTS-Eval and CV3-Eval, outperforming both open-source and leading commercial systems. The inclusion of a "warmed serving protocol" to measure real-time factor (RTF) and first-block latency adds practical relevance to the experimental setup. The comparison against commercial baselines strengthens the claim of production-readiness. However, the specific numerical results are redacted in the provided text, preventing a precise verification of the magnitude of improvement, though the qualitative claims are strong.
The paper provides detailed descriptions of the architecture, including the tokenizer design, the masked diffusion formulation, and the training schedule. The mention of "0.6B backbone lineage" and "1 million hours of speech" provides scale context. However, the specific hyperparameters for the GRPO stage and the exact implementation details of the block-causal attention masking are likely critical for reproduction and may be subject to interpretation. The lack of publicly released code or weights (implied by "none" for URLs) significantly hinders immediate reproducibility for the broader community, although the technical report format suggests a high level of detail.
The primary limitation is the reliance on a specific RVQ tokenizer and the assumption that the semantic anchoring of the first codebook generalizes well across all languages and domains. The block-diffusion approach, while reducing latency, still requires multiple denoising steps per block, which may not be as fast as single-step AR decoding in extreme low-latency scenarios. Furthermore, the complexity of training a diffusion LLM with RL post-training is significantly higher than standard AR TTS training, potentially limiting accessibility. The paper does not extensively discuss the failure modes of the NVV control or the robustness of the emotion conditioning under adversarial text inputs.
This work represents a significant step towards making diffusion-based generative models viable for real-time, production-grade audio applications. By bridging the gap between the quality of diffusion models and the latency requirements of streaming TTS, it expands the toolkit available for developers building voice assistants, gaming NPCs, and accessibility tools. The emphasis on expressive control (emotion, NVVs) aligns with the growing demand for more natural and engaging human-computer interaction. However, the potential for misuse in deepfake generation remains a concern, necessitating robust safety measures in deployment. Luna-TTS introduces a scalable, block-diffusion-based TTS framework that achieves state-of-the-art quality and latency by adapting pretrained LLMs through progressive architectural changes and RL post-training, effectively resolving the latency-quality trade-off inherent in previous diffusion TTS systems.
Spoken dialogue is a natural form of human--computer interaction, yet most speech language models remain limited to turn-based operation and lack real-time adaptability, such as user barge-in. Recent duplex speech-to-speech and speech-to-text models reduce latency by replacing multi-stage pipelines, but often compromise speech quality because accurate ASR, interruption handling, and high-fidelity synthesis must be optimized jointly. We propose VoiceChat-TTS, a low-latency, continuous, and streamable text-to-speech model for interactive agents. VoiceChat-TTS is driven directly by LLM text-token streams, supports explicit interruption via control tokens, and produces silence when no textual input is available. The model enables always-on, responsive speech generation while preserving modularity and high speech quality, and it supports mid-utterance interruptions without resetting the KV cache.
Primary: NVIDIA Corporation
All Institutions: NVIDIA Corporation
VoiceChat-TTS presents a robust, low-latency, and interruptible streaming TTS system that effectively bridges the gap between high-quality offline synthesis and the real-time demands of full-duplex interactive agents. The comprehensive evaluation, including detailed latency analysis and interruption handling benchmarks, demonstrates its practical utility and sets a strong baseline for future research in continuous speech synthesis.
The paper proposes VoiceChat-TTS, a continuous, streamable TTS model built upon the Audio Flamingo 3-Chat architecture. The core technical contributions involve adapting a streaming decoder for full-duplex interaction. Key modifications include: 1) A Character-Aware Subword Encoder to handle out-of-vocabulary subwords from LLMs by converting them to character sequences processed by a shallow Transformer. 2) An interruption mechanism using explicit control tokens to halt generation and transition to silence without resetting the KV cache. 3) Audio Prompt Conditioning using a 3-second reference audio to stabilize speaker identity at the start of generation, addressing the lack of context in streaming settings. 4) A Mixture of Gaussian Estimation Head (MoGH) for faster RVQ token decoding. The approach is pragmatic, focusing on engineering solutions (gated fusion, specific tokenization, silence modeling) to enable real-time, interruptible speech synthesis. While not fundamentally novel in terms of architecture (it extends existing streaming decoders), the specific integration of these components for a unified duplex TTS system is a valuable engineering contribution.
The evaluation is comprehensive and well-structured. The authors compare VoiceChat-TTS against strong offline (Chatterbox-TTS, Qwen3-TTS) and streaming (Audio Flamingo 3-Chat) baselines. Metrics include CER, WER, SECS (speaker similarity), and Squim-MOS. The results show that VoiceChat-TTS achieves competitive quality, significantly outperforming the base Audio Flamingo 3-Chat decoder in intelligibility and quality. The paper also includes a dedicated interruption evaluation using the Full-Duplex-Bench (FDB) subset, measuring Stop Latency and Leakage. The latency analysis is particularly strong, providing detailed breakdowns of acoustic-token ITL and codec decoding time on specific hardware (RTX A6000), demonstrating a 2.1x speedup over a comparable baseline. The use of both synthetic and real conversational data for training is well-justified.
The paper provides significant detail for reproduction. It specifies the model size (977M parameters), the base architecture (Gemma 3-based), the codec configuration (12.5 Hz, 31-codebook RVQ), and the training stages (pretraining on single-turn data, fine-tuning on multi-turn). The code is publicly available via NVIDIA NeMo Speech, and the model checkpoint is on Hugging Face. The training data sources are listed (LibriTTS, HiFiTTS, synthetic data, Fisher). The evaluation protocol is also clearly defined, including the use of Silero VAD for interruption timing detection. This high level of detail ensures that the work is reproducible.
The authors acknowledge several limitations. First, the model lacks user-audio conditioning, meaning it cannot dynamically adapt its prosody to the user's speech characteristics or interruptions in real-time (it relies on text tokens). Second, there is a lack of controlled component-wise ablations; while preliminary experiments suggested cumulative gains, a systematic ablation study is missing. Third, speaker similarity (SECS) degrades over longer sequences of continuous generation, particularly for unseen speakers, indicating a challenge in maintaining long-context speaker consistency. Finally, the evaluation of "silence" generation relies on ASR-based metrics, which may not fully capture the perceptual quality of silence or the naturalness of the transition into/out of silence.
VoiceChat-TTS contributes to the development of more natural and responsive human-computer interaction systems. By enabling low-latency, interruptible speech synthesis, it facilitates the deployment of always-on voice assistants and conversational agents that behave more like human interlocutors. This has positive implications for accessibility, customer service, and companion AI. However, the ease of generating high-quality, realistic speech also raises concerns about potential misuse in deepfakes or deceptive interactions, although the model's reliance on text input mitigates some of the risks associated with direct audio-to-audio generation. VoiceChat-TTS presents a robust, low-latency, and interruptible streaming TTS system that effectively bridges the gap between high-quality offline synthesis and the real-time demands of full-duplex interactive agents. The comprehensive evaluation, including detailed latency analysis and interruption handling benchmarks, demonstrates its practical utility and sets a strong baseline for future research in continuous speech synthesis.
Recent continuous autoregressive TTS models operate directly on continuous speech representations, preserving rich acoustic details while leveraging the instruction-following capabilities of text LLMs. This paradigm opens new possibilities for voice cloning, instruction-controlled voice design, and speech editing, but remains susceptible to error accumulation during autoregressive generation. Existing solutions often require additional semantic modules, multi-stage tokenizer training pipelines, or complex autoregressive architectures. In this work, we propose FireRedTTS3, a simple yet effective speech generation and editing framework that mitigates error accumulation at the representation level. Specifically, we leverage a frozen Audio Encoder trained on diverse speech understanding tasks as a semantic teacher to regularize the audio feature space. This improves text-speech alignment and stabilizes autoregressive generation while keeping the overall system simple. FireRedTTS3 provides two variants: FireRedTTS3-Base for multilingual and multi-dialect zero-shot voice cloning, and FireRedTTS3-Instruct for unified voice cloning, instruction-controlled voice design, and speech editing. Experiments show that FireRedTTS3-Base achieves the best average speech intelligibility and speaker similarity among compared systems on Seed-TTS-Eval and MiniMax-MLS-Test, while FireRedTTS3-Instruct outperforms competing systems on InstructTTSEval and Ming-Freeform-Audio-Edit. These results demonstrate that semantically enriched continuous speech representations, combined with a simple architecture, enable stable, controllable, and high-fidelity speech generation and editing. Code and models are available at https://github.com/FireRedTeam/FireRedTTS3.
Primary: Xiaohongshu
All Institutions: Xiaohongshu
[This paper presents FireRedTTS3, a unified speech generation and editing framework that mitigates error accumulation in continuous autoregressive TTS by using a frozen audio encoder as a semantic teacher to regularize the latent space, achieving state-of-the-art performance in zero-shot cloning and instruction-controlled editing.]
The paper proposes FireRedTTS3, a framework for continuous autoregressive text-to-speech (TTS) that addresses error accumulation through semantically enriched representations. The core innovation is the RedAE tokenizer, which uses a frozen, pretrained Audio Encoder as a semantic teacher to regularize the latent space during training. This approach aims to improve text-speech alignment and stability without adding complex architectural components or multi-stage training pipelines. The generation backbone utilizes an LLM-DiT (Diffusion Transformer) architecture, leveraging the instruction-following capabilities of large language models (Qwen3) for both zero-shot cloning (Base variant) and instruction-controlled editing/design (Instruct variant). The methodology is sound and addresses a known pain point in continuous autoregressive TTS (error accumulation), offering a relatively simple solution compared to prior works that require additional semantic modules or complex tokenizers.
The authors evaluate FireRedTTS3 on four benchmarks: Seed-TTS-Eval, MiniMax-MLS-Test, InstructTTSEval, and Ming-Freeform-Audio-Edit. The results claim state-of-the-art performance in terms of speech intelligibility, speaker similarity, and instruction-following capabilities. The evaluation covers multilingual zero-shot cloning, instruction-controlled voice design, and speech editing. The use of established benchmarks adds credibility. However, the reliance on automated metrics (WER/CER, WavLM similarity, LLM-as-a-judge) for instruction following introduces potential bias, although this is standard in the field. The claim of "best average" performance is supported by the provided tables (referenced as [REF]), suggesting strong empirical results.
The paper provides a GitHub link for code and models. The methodology describes the architecture (Qwen3-style Transformers, DiT, RedAE) and training objectives (GAN, flow-matching, semantic distillation). The use of pretrained components (Qwen3, CAM++, FireRedAudio) aids reproducibility. However, the specific details of the "FireRedAudio" model and the exact composition of the 500k hours of training data are not fully detailed, which may hinder exact replication. The description of the Aggregator and DiT modules is sufficiently detailed for implementation by researchers familiar with these architectures.
The paper acknowledges that Cantonese error rates are high due to the recognition model (Whisper-large-v3) limitations, not necessarily synthesis quality. The reliance on a frozen semantic teacher requires the existence of a high-quality pretrained audio understanding model, which may not be available for all languages or domains. The "simple" architecture still relies on large pretrained LLMs and DiTs, implying significant computational resources for training and inference. The generalization to unseen languages (Portuguese, Ukrainian) is claimed but not deeply analyzed regarding the source of this generalization (linguistic universality vs. acoustic similarity).
FireRedTTS3 contributes to the advancement of controllable and editable speech synthesis, which has applications in accessibility, entertainment, and content creation. The ability to perform instruction-controlled editing and voice design enhances user interaction with speech technologies. However, the ease of voice cloning and editing raises concerns about deepfakes and misuse for fraudulent activities. The open-source release of code and models necessitates careful consideration of safety guardrails and usage policies. [This paper presents FireRedTTS3, a unified speech generation and editing framework that mitigates error accumulation in continuous autoregressive TTS by using a frozen audio encoder as a semantic teacher to regularize the latent space, achieving state-of-the-art performance in zero-shot cloning and instruction-controlled editing.]
Large Audio Language Models (LALMs) have made rapid progress on standardized benchmarks, yet their deployment in practical media workflows, curation, archival indexing, and content distribution remains largely unrealized. We identify automated audio chapterization, the task of segmenting continuous audio streams into thematically coherent chapters, as a demanding and commercially consequential setting that exposes this gap. Chapterization is challenging because boundaries are defined less by objective acoustic events than by subjective editorial judgment, requiring models to reason sequentially over long acoustic contexts and approximate creator-authored boundary decisions. We present AudioChaps, a post-training framework for aligning end-to-end LALMs for this task via Group Relative Policy Optimization (GRPO) guided by Chain-of-Thought (CoT) reasoning. To support training and evaluation, we curate three datasets: AudioChaps-Alignment, derived from creator-annotated chapter boundaries on YouTube; AudioChaps-CoT, which provides structured supervision for well-formatted, high-quality, and evidence-grounded boundary reasoning; and AudioChaps-Eval, a held-out benchmark for audio chapterization. Applying GRPO directly without a Supervised Fine-Tuning (SFT) cold start, AudioChaps-R1-Zero already improves average F1 by 33 points over the state-of-the-art LALM Audio-Flamingo-3-Think. The AudioChaps framework produces our final aligned LALM, AudioChaps-R1, which improves average F1 by 49 points. These results demonstrate that GRPO-trained LALMs can reliably transform unstructured auditory streams into navigable, structured media. Our code, models, and dataset resources will be released upon acceptance at https://github.com/ta012/AudioChaps.
Primary: University of Surrey
All Institutions: University of Surrey, Huawei Noah's Ark Lab
The paper presents a well-executed study on aligning LALMs for audio chapterization using a novel synthetic CoT dataset and GRPO, demonstrating significant performance gains and contributing valuable resources to the field.
The paper proposes AudioChaps, a framework for aligning Large Audio Language Models (LALMs) for automated audio chapterization. The core methodological contribution is a two-stage post-training pipeline: (1) Supervised Fine-Tuning (SFT) on a synthetic Chain-of-Thought (CoT) dataset (AudioChaps-CoT) to establish a structured reasoning format, and (2) Group Relative Policy Optimization (GRPO) to calibrate boundary decisions against creator-authored annotations. The construction of the CoT dataset via an "audio-to-text modality bridge" using a stronger model (Step-Audio-R1) to generate perception logs and reasoning traces is a notable engineering approach to address the lack of reasoning data for subjective audio tasks. The use of GRPO with rule-based rewards (format and accuracy) is a standard application of recent RL techniques but applied effectively to the audio domain. The approach is sound and addresses a specific gap in LALM capabilities (subjective segmentation vs. objective event detection).
The experimental setup is robust. The authors curate three datasets (Alignment, CoT, Eval) stratified across four acoustic regimes (structured speech, dynamic media, gaming, music). They evaluate on both clip-level binary classification and full-length recording chapter detection. The results show significant improvements over the base model (AF3-Think-8B) and a larger baseline (Step-Audio-R1-32B). The ablation studies effectively isolate the contributions of SFT and GRPO, demonstrating that SFT improves recall while GRPO improves precision. The comparison with an ASR-LLM cascade highlights the advantage of end-to-end audio processing for non-speech cues. The inclusion of human evaluation for reasoning quality adds value. The statistical significance testing via bootstrap is a good practice.
The paper provides detailed descriptions of the dataset construction, training prompts, and hyperparameters. The code and models are promised upon acceptance. The use of standard open-source models (AF3-Think, MOSS-Think) as backbones enhances reproducibility. The synthetic data generation pipeline is described clearly enough to be replicated.
The primary limitation is the reliance on synthetic CoT data, which may contain biases or errors from the teacher model. The task is formulated as a binary boundary detection within a window, which may not capture global context as effectively as native long-context models. The evaluation is limited to YouTube content, which may not generalize to all audio types (e.g., pure music without speech, highly noisy environments). The "creator-annotated" ground truth is a proxy for editorial judgment and may vary in quality or consistency.
This work contributes to the practical deployment of LALMs in media workflows, enabling better content navigation and archival indexing. It demonstrates the potential of RL alignment for subjective audio understanding tasks. The released datasets and benchmarks will facilitate further research in audio chapterization and reasoning. The paper presents a well-executed study on aligning LALMs for audio chapterization using a novel synthetic CoT dataset and GRPO, demonstrating significant performance gains and contributing valuable resources to the field.
Semantic speech tokens should preserve linguistic content while suppressing speaker- and duration-dependent variation inherited from acoustic inputs. We propose Iterative Semantic Token Purification (ISTP), an alternating speech-to-unit (S2U) and text-to-unit (T2U) training procedure guided by text predictability. Starting from an initial S2U tokenizer, each iteration trains a T2U model on its deduplicated token sequences. The decoded T2U predictions then serve as connectionist temporal classification targets for a newly initialized S2U model, whose outputs supervise the next T2U model. This cycle progressively aligns the two token generators and biases the token space toward information recoverable from text. Experiments on Mandarin and English show substantially improved S2U--T2U agreement. Independently trained de-tokenizers further show that the refined S2U and T2U tokens retain sufficient content for high-intelligibility voice conversion and text-to-speech synthesis. In voice conversion, the generated speaking rate follows the reference more closely. The refined tokens also exhibit substantially improved cross-speaker consistency and reduced probe-recoverable speaker information.
Primary: City University of Hong Kong
All Institutions: City University of Hong Kong, Huawei Leibniz Research Center, Chinese University of Hong Kong
[One sentence main contribution]. The paper introduces an iterative refinement procedure that aligns speech and text tokenizers to produce speaker-invariant semantic units. [Comprehensive analysis of the technical contribution, methodology, and significance to the field]. This research presents a robust method for enhancing the semantic purity of speech tokens by leveraging the inductive bias of text-to-unit models. By iteratively filtering S2U outputs through a T2U model conditioned only on text, the method effectively strips away speaker and duration variations, resulting in tokens that are more consistent across speakers and languages. The comprehensive evaluation, including independent de-tokenizer testing and speaker probing, strongly validates the approach's effectiveness in balancing content preservation with speaker invariance. This contributes meaningfully to the field of speech tokenization by offering a practical solution to the long-standing challenge of acoustic variability in semantic representations.
The paper proposes Iterative Semantic Token Purification (ISTP), a method to align speech-to-unit (S2U) and text-to-unit (T2U) tokenizers. The core innovation is an alternating training loop: an initial S2U tokenizer generates tokens, which are deduplicated and used to train a T2U model. The T2U model then generates pseudo-targets from text, which are used to train a *newly initialized* S2U model via CTC. This cycle repeats, theoretically forcing the S2U tokenizer to discard speaker-specific and duration-dependent variations that are not recoverable from text. The methodology is logically sound and addresses a genuine problem in speech tokenization: the mismatch between acoustic variability and linguistic invariance. However, the approach is essentially a sophisticated form of self-distillation or knowledge transfer using a text-conditioned model as a "filter." While effective, the conceptual novelty is moderate as it relies on established components (HuBERT, FSQ, BART, CTC) in a novel iterative arrangement.
The experiments are comprehensive and well-designed. The authors evaluate on both Mandarin and English, using multiple metrics: S2U-T2U agreement (WER/BLEU), generation utility (TTS/VC WER/SIM), cross-speaker consistency (UED, SelfBLEU-4), and speaker privacy (probe accuracy). The results show significant improvements in consistency and speaker invariance without degrading content preservation. The use of independent de-tokenizers for evaluation is a strong methodological choice that isolates token quality from decoder bias. The inclusion of speaking rate analysis provides additional insight into duration normalization. The baselines are relevant (StableToken, R-Spin, FlexiCodec, etc.). The results are convincing and support the claims.
The paper provides sufficient detail regarding the architecture (HuBERT, FSQ, BART, CTC) and training data sources (LibriSpeech, GigaSpeech, AISHELL-2, etc.). The iterative procedure is clearly defined. However, specific hyperparameters for the BART model, learning rates, and the exact deduplication logic (beyond "consecutive repetitions") are not fully detailed in the text provided. The claim of "8,000 hours" of training data is substantial, but the reproducibility relies on the availability of the specific SOTA components (DSA-Tokenizer, CA-F5-TTS) and the precise implementation of the iterative loop.
The paper does not explicitly discuss the computational cost of the iterative process, which involves training multiple models from scratch in each iteration. This could be a significant barrier to adoption. Additionally, the method relies on the quality of the initial T2U model; if the initial S2U tokens are too noisy or misaligned, the T2U model might learn incorrect mappings. The paper assumes that "text-predictable" content is the only desirable content, which might discard useful prosodic or emotional cues that are not strictly phonetic but are recoverable from context in some T2U models.
This work has significant implications for privacy-preserving speech technologies and efficient speech generation. By reducing speaker information in tokens, it mitigates privacy risks. The improved consistency aids in building more robust multilingual speech systems. The potential for misuse in creating deepfakes with high linguistic fidelity but anonymized voices is a minor concern, but the primary impact is positive for privacy and efficiency. [One sentence main contribution]. The paper introduces an iterative refinement procedure that aligns speech and text tokenizers to produce speaker-invariant semantic units. [Comprehensive analysis of the technical contribution, methodology, and significance to the field]. This research presents a robust method for enhancing the semantic purity of speech tokens by leveraging the inductive bias of text-to-unit models. By iteratively filtering S2U outputs through a T2U model conditioned only on text, the method effectively strips away speaker and duration variations, resulting in tokens that are more consistent across speakers and languages. The comprehensive evaluation, including independent de-tokenizer testing and speaker probing, strongly validates the approach's effectiveness in balancing content preservation with speaker invariance. This contributes meaningfully to the field of speech tokenization by offering a practical solution to the long-standing challenge of acoustic variability in semantic representations.
Long-paragraph fine-grained audio captioning requires models to recover diverse acoustic facts while avoiding omissions and unsupported details. However, prevailing captioners remain passive one-shot generators: once a detail is overlooked, they cannot identify the evidence gap, query the audio for targeted information, or decide when sufficient evidence has been collected. We formulate this task as active evidence acquisition and introduce Agentic Co-Evolution for Captioning (ACE-Cap). The framework uses multi-turn interaction between a Composer and an Instruct model to form a closed evidence-acquisition loop. A Captioner first produces an initial description. Conditioned on this description and the interaction history, a text-only Composer asks targeted questions about unresolved acoustic attributes, while an audio-conditioned Instruct model provides grounded answers. The Composer then decides when to terminate and synthesizes the accumulated evidence into a final caption. ACE-Cap trains these roles through a unified gold-to-prediction reward derived from fixed, gold-grounded multiple-choice questions and a frozen caption-only judge. For credit assignment in variable-length interactions, LOOP-GRPO replaces the trajectory-wide scalar advantage with span-aligned signals: leave-one-out contributions of individual questions to the accumulated evidence, a quality-cost utility for stopping, and an evidence-preservation utility for final synthesis. Role-wise warm-up followed by alternating Composer and Instruct optimization keeps each update a well-defined single-policy problem while allowing the roles to co-evolve. ACE-Cap thus turns captioning from passive one-shot generation into an adaptive process that learns what evidence to acquire, when to stop, and how to preserve it in a long-paragraph caption.
Primary: Kling Team
All Institutions: Kling Team
ACE-Cap introduces a novel agentic framework for fine-grained audio captioning that leverages active evidence acquisition and a specialized reinforcement learning algorithm (LOOP-GRPO) to improve factual accuracy and detail recovery, representing a significant step forward in adaptive audio understanding systems.
The paper proposes ACE-Cap, a framework that reformulates long-paragraph fine-grained audio captioning as an active evidence acquisition problem. The core innovation lies in the agentic loop involving a text-only Composer and an audio-conditioned Instruct model. The Composer identifies missing acoustic details by querying the Instruct model, which provides grounded answers. This addresses the limitation of passive one-shot captioners. The training methodology introduces LOOP-GRPO (Leave-One-Out Per-turn GRPO), which assigns credit to individual questions based on their counterfactual contribution to the final evidence state, rather than using a trajectory-wide scalar advantage. This is a significant methodological contribution to reinforcement learning for agentic systems, specifically addressing the credit assignment problem in variable-length interaction trajectories. The use of a frozen caption-only judge and gold-grounded multiple-choice questions for reward modeling is a practical approach to aligning the agent with fine-grained factual accuracy.
The experimental evaluation is comprehensive, covering multiple benchmarks (Omni-Cloze, MMAR, MMAU, MMSU). The paper demonstrates that ACE-Cap outperforms strong open-source baselines and remains competitive with proprietary models like GPT-4o Audio and Gemini 2.5 Pro. The ablation studies effectively isolate the contributions of the different training stages (Captioner GRPO, Instruct GRPO, Composer LOOP-GRPO, and alternating co-evolution), showing monotonic improvement. The analysis of the Composer's query behavior and the impact of query budget provides valuable insights into the agent's decision-making process. The results support the claim that active evidence acquisition improves fine-grained captioning accuracy.
The paper provides detailed descriptions of the methodology, including the reward function, the LOOP-GRPO algorithm, and the training pipeline. The use of publicly available models (Qwen2.5-Omni-7B, Qwen3-8B) and datasets (ASID-1M) enhances reproducibility. However, the reliance on a proprietary model (Gemini 3.1 Pro) for generating gold captions and MCQs during the offline stage may introduce some variability, although the core training and evaluation are performed on open-source components. The code is not explicitly mentioned as open-sourced in the text, which is a minor drawback for immediate reproducibility.
The paper acknowledges limitations such as the restriction to audio-only inputs and the fixed caption-as-evaluation protocol. The reliance on a fixed question bank derived from gold captions means the reward signal cannot detect unsupported claims outside the predefined set. Additionally, the alternating optimization strategy, while stabilizing training, may not fully capture the complex interdependencies between the Composer and Instruct models compared to joint optimization, although the authors argue it prevents instability. The performance gap with the strongest proprietary models (e.g., Gemini 3.1 Pro on MMAR) suggests there is still room for improvement in handling very complex audio scenes.
This work contributes to the field of audio-language models by introducing an agentic paradigm for fine-grained understanding. The techniques developed for active evidence acquisition and credit assignment in multi-turn interactions have broader applications in other domains requiring iterative reasoning and information gathering, such as visual question answering or document QA. The emphasis on factual accuracy and evidence preservation aligns with the growing need for reliable and interpretable AI systems. ACE-Cap introduces a novel agentic framework for fine-grained audio captioning that leverages active evidence acquisition and a specialized reinforcement learning algorithm (LOOP-GRPO) to improve factual accuracy and detail recovery, representing a significant step forward in adaptive audio understanding systems.
Existing speech retrieval systems rely on fixed similarity matching and cannot adapt to diverse user intents. We introduce INSPIRE, the first benchmark for instruction-aware speech retrieval, in which natural-language instructions dynamically specify relevance criteria, including semantic content, speaker identity, speaking style, environmental sounds, and their combinations. We evaluate four retrieval paradigms: large audio-language models, cascaded pipelines, self-supervised speech models, and contrastive audio-language models. Our results reveal that no current method robustly handles all retrieval intents. Text-based approaches perform relatively better at semantic retrieval but struggle with paralinguistic attributes, while speech-based models are moderately better at capturing acoustic properties but falter at following instructions. These findings highlight the need for unified architectures capable of instruction-aware speech retrieval.
Primary: National Taiwan University
All Institutions: National Taiwan University, NTU Artificial Intelligence Center of Research Excellence (NTU AI-CoRE)
INSPIRE is the first benchmark for instruction-aware speech retrieval, systematically evaluating the ability of various models to follow natural-language instructions for retrieving speech based on semantic, speaker, style, and environmental attributes, revealing a significant gap in current methods' ability to handle multi-attribute, paralinguistic retrieval.
The paper proposes INSPIRE, a benchmark for instruction-aware speech retrieval. The methodology involves constructing four subsets (DailyTalk, VCTK, Expresso, Synthetic) to test retrieval across semantic, speaker, style, and environmental attributes. The core technical contribution is the formalization of the problem and the construction of the dataset, particularly the synthetic subset which uses GPT-4o-mini for TTS and GPT-5.2 for instruction generation. The evaluation methodology involves testing four distinct retrieval paradigms: Large Audio-Language Models (LALMs), Cascaded Pipelines (ASR/Captioning + Text Retrieval), Self-Supervised Speech Models, and Contrastive Audio-Language Models. The approach is sound in its comparative design, aiming to highlight the gaps in current models. However, the novelty is somewhat limited by the fact that it is primarily a benchmarking paper rather than proposing a new architecture or algorithm. The use of proprietary models (GPT-4o-mini, GPT-5.2) for data generation raises questions about the independence of the benchmark from specific model biases, although this is common in modern AI research.
The experimental evaluation is comprehensive in scope, covering a wide range of existing models and retrieval strategies. The results clearly demonstrate the "modality specialization gap": text-based methods excel at semantic retrieval but fail at paralinguistic tasks, while speech-based models do the opposite. The inclusion of ablation studies on instruction types, pooling strategies, and layer-wise analysis adds depth. The use of Recall@K and NDCG is standard and appropriate. The results are robustly presented with clear tables and radar plots. The finding that no current method handles all intents robustly is significant and well-supported by the data. The synthetic subset's quality assessment (WER, speaker accuracy, emotion accuracy) provides confidence in the dataset's validity.
The paper provides a GitHub link to the code and data, which is a strong point for reproducibility. The dataset construction process is described in detail, including the sources (DailyTalk, VCTK, Expresso) and the generation pipeline for the synthetic subset. The baseline implementations are standard and well-documented. The use of open-source models for most baselines (HuBERT, WavLM, CLAP, E5-Mistral, etc.) ensures that other researchers can reproduce the experiments. The reliance on proprietary models for some baselines (Gemini, GPT-4o-mini) limits full reproducibility of those specific comparisons but does not invalidate the overall benchmark.
A key limitation is the reliance on synthetic data for the most complex multi-attribute tasks. While the synthetic data is high-quality, it may not fully capture the nuances and complexities of real-world speech, such as overlapping speakers, extreme noise, or diverse accents. The instruction generation using GPT-5.2 (a hypothetical or very recent model name, possibly a typo for GPT-4o or similar) introduces a potential bias if the instructions are not diverse enough or if they reflect the specific linguistic patterns of the generator. Additionally, the benchmark is static; it does not address the dynamic nature of user queries in real-world applications. The performance of LALMs is still relatively low, suggesting that the field is in its early stages, which is a limitation of the current state-of-the-art rather than the paper itself, but it highlights the difficulty of the task.
This paper has significant broader impact by defining a new and challenging task in speech processing. It highlights the limitations of current multimodal models and provides a clear direction for future research towards unified, instruction-aware speech retrieval systems. This could lead to more intuitive and powerful speech search engines, accessibility tools, and archival systems. The benchmark will likely spur further research in this area, fostering competition and innovation. The emphasis on paralinguistic attributes (style, emotion, speaker) is particularly important for applications requiring nuanced understanding of human communication. INSPIRE is the first benchmark for instruction-aware speech retrieval, systematically evaluating the ability of various models to follow natural-language instructions for retrieving speech based on semantic, speaker, style, and environmental attributes, revealing a significant gap in current methods' ability to handle multi-attribute, paralinguistic retrieval.
Generating personalized dance videos from a reference image, text prompt, and audio track requires music-conditioned body motion. Singing-and-dancing adds a second requirement: the visible subject must also articulate the vocals. Existing music-conditioned methods focus primarily on choreography, while speech-driven models generally assume that the visible subject produces the input voice, leaving this combined setting largely underexplored. We introduce SingDance, a unified video diffusion framework that formulates controllable vocal articulation as a semantic role: the visible subject is either the source, who produces the vocal signal, or the listener, who receives it from an off-screen performer. Hard-compact routing selects task-relevant speech, music, and role conditions, which are composed through frame-wise joint audio injection; source and listener retain the same speech pathway. Training uses asymmetric supervision: on-screen speaking and curated off-screen conversational-response videos establish role control, while instrumental and song-based dancing-only videos establish music-conditioned body motion. The target Song/Source configuration is never observed during training. At inference, assigning the source role to a song composes separately learned articulation and song-conditioned dance capabilities, enabling compositional zero-shot singing-and-dancing. Experiments demonstrate strong motion--beat alignment and visual fidelity, reliable paired switching of vocal articulation while preserving music-aligned body motion, and highly competitive lip synchronization with substantially fewer generation-time parameters than the strongest speech-driven baseline evaluated.
Primary: Kuaishou Technology
All Institutions: Kuaishou Technology
SingDance presents a novel and effective framework for compositional zero-shot singing-and-dancing video generation by introducing role-aware audio conditioning and asymmetric staged training, achieving strong performance in motion-beat alignment and lip synchronization while maintaining high visual fidelity.
The paper proposes SingDance, a video diffusion framework that addresses the compositional zero-shot problem of generating singing-and-dancing videos. The core methodological contribution is the formulation of vocal articulation as a semantic "role" (source vs. listener) rather than a simple binary lip-sync instruction. This allows the model to distinguish between a subject producing vocals and a subject listening to off-screen vocals. The architecture utilizes hard-compact routing to select relevant audio tokens (speech from Wav2Vec 2.0, music from MuQ) and injects them via frame-wise joint audio injection into a Video DiT backbone (Wan2.2). A key innovation is the staged training strategy with asymmetric supervision: role control is learned from speech data, and music-conditioned dance is learned from instrumental/dancing-only data, without ever seeing paired singing-and-dancing data during training. The zero-shot composition at inference (Song/Source) combines these separately learned capabilities. The approach is technically sound and elegantly solves the conditioning conflict between speech and music in a unified framework.
The experimental evaluation is comprehensive. The authors curate specific test sets (SingDance-50, Dance-100) to evaluate the proposed capabilities. They compare against strong baselines including MusicInfuser and Wan-S2V. The results demonstrate strong motion-beat alignment and visual fidelity. Crucially, the paired vocal-role switching experiment effectively verifies that the model can switch between singing and listening while maintaining consistent body motion and audio conditioning, a capability that baseline models lacking explicit role conditioning fail to achieve robustly. The lip synchronization metrics are competitive with specialized speech-driven models, despite the parameter efficiency. The ablation studies support the necessity of the music pathway for beat alignment.
The paper provides detailed implementation details, including the backbone (Wan2.2-TI2V-5B), training hardware (64 A100s), hyperparameters (learning rate, batch size), and specific audio encoders (Wav2Vec 2.0, MuQ). The data sources are described (HuMoSet, MA-Data, proprietary collections), though the proprietary nature of some data may limit full reproducibility of the training set. The project page URL is provided, which likely contains code or weights, enhancing reproducibility potential.
The current model is limited to short, single-person clips with a binary, clip-level vocal role. It does not handle within-clip role transitions, duets, or multiple vocal roles. The reliance on hard-compact routing means the model must be explicitly told the role; it does not automatically infer the role from the audio content, which is a design choice but also a limitation for fully autonomous generation. The use of proprietary data for training is a constraint for independent verification.
This work significantly advances the field of multimodal video generation, particularly in the niche of music and speech-conditioned human animation. By enabling compositional zero-shot singing-and-dancing, it lowers the barrier for creating high-quality, personalized performance videos. The role-aware conditioning framework could be extended to other interactive scenarios. However, the ability to generate realistic singing and dancing videos also raises concerns regarding deepfake misuse, although the explicit role conditioning might offer some avenue for detection or control. SingDance presents a novel and effective framework for compositional zero-shot singing-and-dancing video generation by introducing role-aware audio conditioning and asymmetric staged training, achieving strong performance in motion-beat alignment and lip synchronization while maintaining high visual fidelity.
Automatic video dubbing in the wild remains fundamentally limited by two competing constraints: hierarchical methods depend on brittle, multi-stage preprocessing pipelines that severely restrict data scalability and practical deployment, while holistic approaches operating on uncropped video suffer from weak temporal alignment and speaker-utterance ambiguity in multi-speaker settings. To overcome these limitations, we propose CineDub, a unified diffusion-based model that achieves precise multi-speaker dialogue dubbing directly from uncropped videos, without face cropping or speaker diarization. Central to our approach is the Implicitly-Coupled Holistic Conditioning (ICHC) paradigm, where holistic visual representations and a semantic-bundled transcription format are encoded independently, yet implicitly coupled through cross-modal training to resolve speaker ambiguity and enable precise multi-speaker multi-turn dialogue dubbing. Building on the unified temporal cues captured by holistic visual features, we further extend CineDub to joint speech and audio generation. We introduce an Ambient-to-Linguistic Curriculum Learning (ALC) to mitigate sub-task degradation, and a decoupled textual branch control mechanism to resolve cross-prompt interference during simultaneous generation. We also release two in-the-wild benchmarks, CineDub-Multi for multi-speaker dialogue dubbing and CineDub-SA for video-to-speech-and-audio (V2SA) generation, to enable evaluation under realistic conditions. Experiments show that CineDub achieves state-of-the-art results on established single-speaker dubbing and video-to-audio benchmarks while excelling in multi-speaker dialogue dubbing and acoustically coherent joint generation.
Primary: Monash University
All Institutions: Monash University, University of Chinese Academy of Sciences, Tsinghua University
CineDub presents a robust and innovative approach to multi-speaker video dubbing by effectively decoupling holistic visual conditioning from semantic transcription, achieving state-of-the-art results and providing valuable new benchmarks for the community.
The paper proposes CineDub, a unified diffusion-based framework for end-to-end video dubbing that operates on uncropped videos, addressing the limitations of hierarchical methods (which rely on brittle preprocessing like face cropping and diarization) and holistic methods (which suffer from speaker-utterance ambiguity). The core technical contribution is the Implicitly-Coupled Holistic Conditioning (ICHC) paradigm. This involves using SynchFormer features as a holistic visual condition, which the authors argue captures both event-level audio-visual correspondence and fine-grained lip-sync cues through emergent attention mechanisms. To resolve speaker ambiguity, they introduce a "semantic-bundled transcription" format, where speaker descriptions are coupled with transcript segments, encoded by a pre-trained LLM (Gemma-T5). The framework extends to joint speech and audio generation (V2SA), introducing two key mechanisms: Ambient-to-Linguistic Curriculum Learning (ALC) to mitigate sub-task degradation by training on audio first, then speech, and a decoupled textual branch control mechanism to prevent cross-prompt interference between speech and audio conditions in the diffusion transformer. The approach is technically sound, leveraging recent advances in diffusion transformers and multimodal alignment, but the novelty lies primarily in the specific conditioning strategies and curriculum learning design rather than fundamental architectural changes.
The authors evaluate CineDub on single-speaker dubbing (GRID, CHEM), multi-speaker dubbing (new CineDub-Multi benchmark), video-to-audio (VGGSound), and joint V2SA (new CineDub-SA benchmark). Results show CineDub outperforms hierarchical baselines (like HPMDubbing, Speak2Dub) and holistic baselines (DeepDubber, DeepAudio) on most metrics, particularly in zero-shot voice cloning and multi-speaker scenarios. The introduction of two new benchmarks, CineDub-Multi and CineDub-SA, is a significant contribution, addressing the lack of realistic, in-the-wild evaluation data for multi-speaker and joint generation tasks. The ablation studies effectively demonstrate the necessity of the semantic-bundled transcription, the ALC curriculum, and the decoupled branches. The use of standard metrics (WER, LSE-D, UTMOS, FDVGG) is appropriate, though the reliance on embedding-based metrics for audio quality (FDVGG, KL) has known limitations regarding perceptual fidelity, which the authors partially address with UTMOS.
The paper provides detailed descriptions of the model architecture, training stages, and data processing pipelines. It mentions the use of pre-trained models (SynchFormer, Gemma-T5, CLIP) and specific datasets (VGGSound, AudioSet, SpeakerVid-5M). The release of the CineDub-Multi and CineDub-SA benchmarks enhances reproducibility for future work. However, the code is not explicitly linked in the provided text (only a demo page is mentioned), and the specific hyperparameters for the curriculum learning and meta-token initialization are not fully detailed in the excerpt, which might hinder exact replication.
The paper acknowledges that SynchFormer's attention-switching is not perfectly reliable, occasionally drifting or oscillating, which the semantic-bundled transcription aims to mitigate but may not fully resolve in all edge cases (e.g., heavy occlusion or off-screen speech). The reliance on MLLMs (Gemini 2.5 Pro) for generating the semantic-bundled transcriptions introduces a dependency on external models and potential errors in annotation, although manual verification is claimed. The joint generation model, while competitive, may still suffer from subtle acoustic incoherence compared to specialized single-task models, as indicated by some metric gaps. The benchmarks, while novel, are limited in size (139 and 562 samples respectively), which may not fully capture the diversity of in-the-wild scenarios.
CineDub has significant potential for multimedia production, enabling scalable and realistic video dubbing for movies, TV shows, and online content, particularly in multi-speaker settings. The release of new benchmarks will facilitate further research in this area. However, the technology also raises concerns about deepfake creation and misuse in generating deceptive audio-visual content. The authors' emphasis on realistic evaluation and the complexity of the pipeline may act as a slight barrier to malicious use, but the dual-use nature of such generative models remains a concern. CineDub presents a robust and innovative approach to multi-speaker video dubbing by effectively decoupling holistic visual conditioning from semantic transcription, achieving state-of-the-art results and providing valuable new benchmarks for the community.
Expressive text-to-speech (TTS) systems that use explicit conditioning labels provide direct and interpretable control over expressive attributes, in contrast to reference-based or prompting-based approaches, but require labeled data. Obtaining these labels at scale is costly and time-consuming, yet no prior semi-supervised framework addresses this specific bottleneck. Existing semi-supervised TTS methods instead target scarcity of paired speech-text data or transcriptions. To address the scarcity of expressive labels, we propose an Iterative Self-Learning (ISL) framework for expressive TTS, built on Invert-Classify, a classifier-free method that recovers discrete expressive labels by inverting a frozen generative model. The framework iteratively pseudo-labels unlabeled speech using the current model, retrains on the combined labeled and pseudo-labeled data, and repeats, progressively refining label quality and synthesis. We validate on two expressive tasks, word-level prominence and utterance-level emotion, across multiple low-resource data splits. We find that iterative refinement can improve pseudo-label accuracy over single-pass baselines. Furthermore, we observe that these improvements in pseudo-labeling of expressivity translate to gains in expressive label adherence and synthesis quality, confirmed by objective metrics and human listening tests. In the most data-scarce conditions, ISL-trained models outperform single-pass pseudo-labeling and further approach fully supervised performance, demonstrating that gradient-based ISL is an effective solution to expressive label scarcity in low-resource TTS.
Primary: University of Edinburgh
All Institutions: University of Edinburgh, Huawei, Wallenberg AI, Autonomous Systems and Software Program (WASP)
This paper presents a robust iterative self-learning framework for expressive TTS that effectively leverages unlabeled data to improve pseudo-label quality and synthesis performance in low-resource settings, demonstrating significant gains in both objective metrics and subjective listener preference.
The paper proposes an Iterative Self-Learning (ISL) framework to address the scarcity of expressive labels in Text-to-Speech (TTS). The core methodological contribution is the integration of "Invert-Classify" into an iterative loop. Invert-Classify is a classifier-free technique that recovers discrete expressive labels (emotion, prominence) by optimizing input embeddings via gradient descent on a frozen generative model's loss function. The authors adapt this for flow-matching models (Matcha-TTS) by fixing stochastic sampling parameters during inversion to ensure convergence. The iterative process involves pseudo-labeling unlabeled data, retraining on the combined dataset, and repeating. The methodology is sound and addresses a specific gap in semi-supervised TTS, moving beyond simple pseudo-labeling to a refinement loop. However, the novelty is somewhat limited by the reliance on existing inversion techniques and the standard IPL protocol, adapted specifically for expressive TTS.
The evaluation is comprehensive, covering two distinct tasks (word-level prominence and utterance-level emotion) and multiple low-resource data splits (0.5% to 20%). The authors employ rigorous metrics including TTSDS2, emotion recognition F1, and subjective A/B listening tests using Bradley-Terry models. The results demonstrate that ISL improves pseudo-label accuracy and downstream synthesis quality compared to single-pass baselines, particularly in low-resource regimes. The analysis of training duration (3 vs 10 vs 50 epochs) provides valuable insight into the stability of the self-learning loop, identifying a "saturation" failure mode with excessive retraining. The use of human listening tests adds significant weight to the claims of perceptual improvement.
The paper provides detailed implementation details, including hyperparameters for optimization, inversion steps, and dataset splits. The use of standard backbones (Matcha-TTS, HiFi-GAN) and public datasets (ESD, Naver-Prosody) enhances reproducibility. The description of the inversion process, including the fixing of timesteps and noise tensors, is sufficiently detailed for replication.
The primary limitation is the reliance on a single backbone architecture (Matcha-TTS), although the authors acknowledge this. The method is sensitive to the amount of seed data; too little prevents bootstrapping, while too much yields diminishing returns. The approach assumes that the generative model can effectively invert to recover labels, which may not hold for all expressive traits or model architectures. Furthermore, the computational cost of iterative gradient-based inversion is non-trivial.
This work contributes to the democratization of expressive TTS by reducing the dependency on expensive, manually annotated expressive datasets. It enables the use of large-scale unlabeled speech corpora for training controllable TTS systems, which has broad implications for accessible speech synthesis, virtual assistants, and content creation. The insights into iterative self-learning dynamics are also applicable to other semi-supervised speech tasks. This paper presents a robust iterative self-learning framework for expressive TTS that effectively leverages unlabeled data to improve pseudo-label quality and synthesis performance in low-resource settings, demonstrating significant gains in both objective metrics and subjective listener preference.
Text-to-audio-video (T2AV) generation models produce a video and its soundtrack from a textual description, but offer no control over whose voice speaks in the output. We show that a base T2AV model can be turned into a voice-cloning model by adding a single zero-initialized linear layer on top of its audio backbone, fine-tuning for a comparatively short training schedule, and conditioning on a short reference recording at inference time. The reference is injected through two complementary signals: its diffusion latents are prepended to the audio stream, and a global speaker embedding modulates token of the target audio. On a benchmark of 674 speaker-text pairs spanning 30 speakers we compare against five strong voice-cloning text-to-speech baselines: our enhanced 5B model attains the highest speaker-encoder cosine similarity (SECS) across three independent verification networks (ECAPA-TDNN, WavLM-SV, Resemblyzer), statistically significantly outperforming every baseline. A side product of the architecture is that the audio path can be evaluated without the video path at inference time, yielding a ~30x speed-up over the full audio-video diffusion loop while preserving the voice-cloning behaviour.
Primary: Kandinsky Lab
All Institutions: Kandinsky Lab
The paper presents a practical and effective method for adding voice cloning to T2AV models with minimal architectural changes, achieving state-of-the-art speaker similarity at the cost of some transcription accuracy.
The paper proposes a method to add voice cloning capabilities to existing Text-to-Audio-Video (T2AV) diffusion models. The core innovation is architectural minimalism: adding a single zero-initialized linear layer to inject a global speaker embedding (via FiLM) and prepending reference audio latents to the input sequence. This allows the model to condition on a reference voice without retraining the entire backbone from scratch. The approach leverages the existing self-attention mechanisms for latent prepending and introduces a new modulation path for the speaker embedding. The methodology is sound and builds logically on existing practices in TTS (speaker embeddings) and diffusion (latent conditioning), but the combination within a T2AV context is a novel application. The zero-initialization ensures stability during the fine-tuning phase.
The evaluation is conducted on a benchmark of 674 speaker-text pairs across 30 speakers from the VCTK corpus. The authors compare their method against five strong baselines, including dedicated TTS models (Qwen3-TTS, XTTS-v2, IndexTTS2) and a joint audio-video model (NAVA). Results show that the proposed method achieves higher speaker-encoder cosine similarity (SECS) across three verification networks (ECAPA-TDNN, WavLM-SV, Resemblyzer) compared to baselines. However, this comes at the cost of higher Word Error Rates (WER), indicating a trade-off between speaker fidelity and transcription accuracy. The paper also demonstrates that the audio-only path can be run independently for a ~30x speed-up with minimal loss in speaker similarity. The "no regression" analysis shows that fine-tuning does not degrade the base model's performance on text-to-audio generation.
The paper provides detailed implementation details, including hyperparameters (learning rates, optimizer settings, diffusion steps), architecture dimensions, and training schedules. The use of open-source components (VCTK, Qwen3-TTS encoder, ECAPA-TDNN, etc.) aids reproducibility. However, the base T2AV model (Kandinsky 5.0) is described as having an "internal corpus" and the specific checkpoints are not publicly linked in the text provided, which may hinder exact replication. The code is not explicitly linked in the provided text.
The primary limitation is the trade-off between speaker similarity and speech intelligibility (WER). The model tends to preserve acoustic quirks of the reference, which can lead to hallucinations or errors in the generated text, especially on short or difficult prompts. The performance is also sensitive to reference length and language matching. The method relies on the availability of a pre-trained T2AV model, limiting its applicability to models that have not been trained on such data.
This work enables more personalized and controllable audio-visual generation, which has applications in content creation, dubbing, and virtual avatars. However, the ability to clone voices raises significant ethical concerns regarding misuse for deepfakes or non-consensual voice synthesis. The authors should have included a more robust discussion on these risks and potential mitigation strategies. The paper presents a practical and effective method for adding voice cloning to T2AV models with minimal architectural changes, achieving state-of-the-art speaker similarity at the cost of some transcription accuracy.
Large audio-language models (LALMs) make it possible to interact with language models through speech, music, and environmental sound, but they also introduce a safety surface that is difficult to expose with text-only red-teaming. We study automated audio-grounded red-teaming, where a text query must remain safe in isolation while the joint text-audio input induces harmful target behavior. We propose ARENA, a closed-loop framework that trains a controller on an independent 2,000case text-audio dataset. MD-Judge supplies training rewards and adaptive search feedback, while a separate, non-adaptive Llama Guard 3 evaluator alone labels final outcomes. On 520 held-out AdvBench objectives, ARENA achieves FDR/PSR of 87.9/100.0%, 71.5/96.3%, 68.1/100.0%, and 75.4/98.5% on Audio Flamingo 3, Qwen2-Audio, MiMo-Audio, and GPTAudio, respectively. Ablations show that feedback-based refinement and audio-variant search substantially improve attack discovery.
Primary: Unknown
All Institutions: Unknown
[One sentence main contribution]. [The paper presents ARENA, a novel automated red-teaming framework that uses closed-loop feedback and preference optimization to generate effective audio-grounded jailbreaks for Large Audio-Language Models, demonstrating significant vulnerabilities in current safety mechanisms across multiple state-of-the-art models.]
The paper proposes ARENA, a closed-loop automated red-teaming framework for Large Audio-Language Models (LALMs). The core methodology involves training a controller LLM to generate text-safe queries paired with audio prompts (either speech or environmental sound) that induce harmful compliance in target models. The training utilizes a 2,000-case seed pool, employing Reward-Weighted Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) based on feedback from an MD-Judge model. During inference, the controller iteratively refines prompts based on judge feedback until a success threshold is met. A key technical distinction is the separation of the search feedback mechanism (MD-Judge) from the final evaluation metric (Llama Guard 3), which is intended to prevent overfitting to the judge. The approach addresses the specific challenge of "audio-grounded" jailbreaks where safety is conditional on the audio modality.
The authors evaluate ARENA on four LALMs: Audio Flamingo 3, Qwen2-Audio, MiMo-Audio, and GPT-Audio. They use 520 held-out AdvBench objectives. The results show high Fault Detection Rates (FDR), ranging from 68.1% to 87.9%, significantly outperforming static baselines like AJailBench and JALMBench. The paper includes ablation studies on refinement budgets, target sampling parameters, and audio variant counts, demonstrating that feedback-based refinement and audio variation substantially improve attack discovery. Transferability analysis shows that attacks found on one model can partially transfer to others. The evaluation is comprehensive, covering multiple models and providing detailed failure analysis.
The paper provides a GitHub link for code. The methodology describes the training data construction (2,000 seeds), the reward shaping formulas, and the DPO setup. However, the specific versions of the TTS models (Piper, TangoFlux) and the exact configuration of the MD-Judge and Llama Guard 3 evaluators are not fully detailed in the text provided, which may hinder exact reproduction. The separation of training and evaluation data is clearly stated, which is good practice.
The primary limitation is the reliance on LLM-based judges (MD-Judge and Llama Guard 3) for both training feedback and final evaluation. While the separation is intended to mitigate this, LLM judges are known to have biases and inconsistencies, particularly with audio-grounded reasoning. The paper does not provide human evaluation of the generated jailbreaks or the harmfulness of the responses, which is a significant gap for safety research. Additionally, the effectiveness is limited to the specific audio synthesis models used; different TTS or audio generation models might yield different results. The focus on automated red-teaming also means it may miss nuanced social engineering attacks that require more complex, multi-turn human interaction.
This work has significant implications for the safety of multimodal AI systems. By exposing vulnerabilities in LALMs, it helps developers identify and patch safety gaps before deployment. However, the release of such a powerful red-teaming tool also raises dual-use concerns, as the generated jailbreaks could be misused by malicious actors. The paper responsibly frames this as a safety auditing tool, but the availability of the code and methodology requires careful consideration of access controls. [One sentence main contribution]. [The paper presents ARENA, a novel automated red-teaming framework that uses closed-loop feedback and preference optimization to generate effective audio-grounded jailbreaks for Large Audio-Language Models, demonstrating significant vulnerabilities in current safety mechanisms across multiple state-of-the-art models.]
Audio-Text Foundation Models (ATMs) fail catastrophically under severe acoustic noise, yet existing adaptation strategies either rely on gradient-based Test-Time Adaptation (TTA), which reinforces noise rather than signal, or on prompt tuning that requires privileged noise annotations unavailable at inference. We address these failures with PRISM (Prototype-Rectified Iterative Self-supervised Manifold Denoising), a training-free, source-free TTA framework grounded in the Affine Noise Hypothesis: severe acoustic noise induces a low-rank affine shift in the multimodal latent space, with more than 90% of distortion energy confined to the leading 60 principal components. PRISM estimates and reverses this distortion from an unlabeled target batch using frozen text prototypes as geometric anchors via three closed-form geometric corrections compiled into a single static projection matrix by Affine Bias Regression. At inference, adaptation reduces to one matrix-vector multiplication in 0.0009 ms, making it substantially faster than gradient-based TTA while requiring no additional training. On UrbanSound8K, PRISM improves over the zero-shot baseline by 12.94 percentage points and surpasses an oracle-assisted TTA baseline by 9.41 percentage points, despite never observing its privileged augmented noise prompts. We further identify the Polyphonic Trap, a principled failure mode of subspace deflation for broadband classes, and resolve it via Confidence-Aware Regression (CAR), recovering up to 8.16 percentage points for the worst-affected class.
Primary: Indian Institute of Science Education and Research Bhopal
All Institutions: Indian Institute of Science Education and Research Bhopal, Vellore Institute of Technology Bhopal
This paper presents a novel, geometric approach to test-time adaptation for audio-text models, demonstrating that affine corrections in latent space can effectively mitigate severe acoustic noise without gradients or source data.
The paper proposes PRISM, a training-free, source-free Test-Time Adaptation (TTA) framework for Audio-Text Foundation Models (ATMs). The core theoretical contribution is the "Affine Noise Hypothesis," which posits that severe acoustic noise induces a low-rank affine shift in the multimodal latent space. To correct this, PRISM employs three closed-form geometric operations: Orthogonal Procrustes Cross-modal Alignment (OPCA) to align manifolds, Class-Conditioned Variance Deflation (CCVD) to remove noise-dominant directions via Fisher Linear Discriminant Analysis, and Per-Class Residual Translation. These are compiled into a static projection matrix via Affine Bias Regression (ABR). The approach is mathematically grounded in linear algebra and manifold learning. While the geometric intuition is sound, the novelty lies primarily in the specific combination and application of these techniques to the ATM domain, rather than the introduction of fundamentally new mathematical primitives. The "Polyphonic Trap" analysis is a valuable diagnostic contribution, identifying a specific failure mode where high within-class variance in polyphonic sounds is mistaken for noise.
The evaluation is conducted on UrbanSound8K, ESC-50, and DCASE/TAU 2019 datasets with injected noise. The results show significant improvements over zero-shot baselines and other TTA methods like PCA++ and TDA. Notably, PRISM outperforms the oracle-assisted ContextDA baseline, which is a strong result given that ContextDA has access to privileged noise annotations. The paper provides detailed ablation studies and sensitivity analyses. However, the evaluation relies heavily on synthetic noise injection. While the SNR sweep is thorough, the generalization to real-world, non-stationary acoustic environments (beyond the TAU corpus) is less rigorously demonstrated. The comparison with gradient-based TTA methods highlights the speed advantage but does not fully explore the accuracy-latency trade-off in dynamic streaming scenarios where batch sizes might be smaller than the calibration buffer.
The paper provides detailed algorithmic steps, hyperparameters (K=60, p=0.8, etc.), and implementation details (LAION-CLAP checkpoint). The closed-form nature of the solution enhances reproducibility. However, the code is not publicly linked in the provided text, and the specific prompt templates used for text prototypes are only partially described ("20 diverse prompt templates"). The reliance on a specific foundation model (CLAP) limits direct generalizability to other architectures without adaptation.
The primary limitation is the "Polyphonic Trap," where the method fails for broadband, spectrally dense classes like street music. Although the authors propose Confidence-Aware Regression (CAR) to mitigate this, it adds complexity and the method still struggles with classes where semantic variance overlaps with noise subspace geometry. Additionally, the method assumes the noise distortion is low-rank and affine; if the acoustic environment induces complex, non-linear manifold warping that violates this hypothesis, performance may degrade. The method also requires a calibration batch, which may not be feasible in strictly real-time, single-sample inference scenarios without a warm-up period.
This work contributes to the robustness of audio foundation models in real-world, noisy environments, which is critical for applications like assistive listening, environmental monitoring, and mobile audio search. By providing a computationally efficient, training-free solution, it lowers the barrier for deploying robust ATMs on edge devices. The identification of the Polyphonic Trap offers insights into the limitations of subspace-based denoising for complex audio classes, guiding future research in geometric deep learning for audio. This paper presents a novel, geometric approach to test-time adaptation for audio-text models, demonstrating that affine corrections in latent space can effectively mitigate severe acoustic noise without gradients or source data.
We address the problem of encoding room impulse responses (RIRs) into high-order Ambisonics (HOA) representations from arbitrary and potentially insufficient or incomplete microphone array measurements. This task is fundamentally ill-posed for microphone arrays with limited spatial capture capabilities, such as irregular or sparse arrays, as classical linear methods fail to reconstruct high-order spatial detail. We introduce a diffusion-based generative framework that models the statistical properties of HOA RIRs. This enables device-agnostic encoding from arbitrary microphone arrays, potentially unseen during data measurement. Our approach incorporates a posterior sampling procedure that enforces consistency between the estimated signals and the measurements while plausibly reconstructing spatial information that is unobservable from the limited measurements alone. Experiments on simulated data demonstrate that our method outperforms linear and neural baselines, achieving accurate HOA RIR estimation up to 12th order. A listening test with binaural renderings, including both simulated and measured RIRs, further confirms that the proposed method yields higher perceptual similarity to reference Ambisonics RIRs than all baselines. The flexibility and accuracy of the proposed framework opens new possibilities for scalable acoustics simulations.
Primary: Reality Labs Research, Meta
All Institutions: Reality Labs Research, Meta, Aalto University, Acoustics Lab
This paper presents a significant advancement in spatial audio processing by effectively leveraging diffusion models to solve the ill-posed problem of high-order Ambisonics encoding from sparse measurements, achieving state-of-the-art perceptual and objective performance.
The paper proposes a novel application of diffusion models to the ill-posed inverse problem of encoding Room Impulse Responses (RIRs) into High-Order Ambisonics (HOA) from sparse or irregular microphone arrays. The core methodological contribution is the integration of a device-agnostic diffusion prior with a posterior sampling procedure that enforces data consistency via a range-projected likelihood term. The use of a hybrid time-frequency/time-domain architecture (NCSN++ backbone with a U-Net refinement stage for early reflections) is a well-reasoned design choice given the distinct temporal characteristics of RIRs. The introduction of a compressed spectrogram distance metric for the likelihood guidance is a technically sound innovation to emphasize weak high-order components. The formulation separates the device-agnostic prior from the device-specific likelihood, which is a strong theoretical foundation for generalization.
The experimental evaluation is rigorous and comprehensive. The authors utilize two datasets: a large-scale internal FDTD simulation dataset and the public Treble-10 dataset. They compare against three strong baselines: Linear Least-Squares, a Time-Dependent Neural Encoder, and a Conditional Diffusion model specific to the device. The inclusion of both objective metrics (EDC, NPM) and a subjective listening test (MUSHRA-style with binaural rendering) provides robust validation. The results demonstrate clear superiority in both perceptual similarity and objective metrics, particularly in preserving early reflections and late reverberation tails. The ablation studies effectively isolate the contribution of the range-projection constraint and the hybrid architecture.
The paper provides sufficient detail for reproduction, including dataset descriptions, model architectures (NCSN++ based), training hyperparameters (AdamW, learning rate, batch size), and specific implementation details like the compressed spectrogram definition. The use of standard libraries (PyTorch) and well-known architectures aids reproducibility. However, the reliance on an internal FDTD dataset for the primary training results limits independent verification of the scale of the results, though the public Treble-10 results offer some ground truth.
The primary limitation is the distribution gap between simulated training data and real-world measured RIRs, which the authors acknowledge and demonstrate leads to performance drops on measured data (Eigenmike-64). The method assumes a highly accurate Array Transfer Function (ATF), which may not hold in real-world deployments with calibration errors. Additionally, the iterative sampling process is computationally intensive, making it less suitable for real-time applications without significant acceleration techniques. The evaluation is currently limited to a single device configuration (Aria Glasses) for the device-agnostic claim, though the method is theoretically general.
This work has significant implications for spatial audio processing, particularly for Virtual Reality (VR) and Augmented Reality (AR) applications where scalable and accurate acoustic simulation is crucial. By enabling high-quality HOA encoding from sparse arrays, it lowers the barrier for capturing spatial audio with wearable devices. It also facilitates the generation of large-scale training data for spatial audio models. The device-agnostic nature of the approach promotes interoperability across different hardware platforms. This paper presents a significant advancement in spatial audio processing by effectively leveraging diffusion models to solve the ill-posed problem of high-order Ambisonics encoding from sparse measurements, achieving state-of-the-art perceptual and objective performance.
Music foundation models are commonly used as frozen audio feature extractors, yet selecting which layer to extract from remains largely heuristic. Current practice defaults to fixed depths or multi-layer fusion, with limited understanding of why certain layers transfer better across downstream tasks or how representation quality varies with depth and pre-training paradigm. We conduct a systematic layer-wise analysis of 12 music foundation models spanning three pre-training paradigms (masked modeling, autoregressive modeling, and contrastive learning), characterizing their hidden representations through intrinsic geometric and transformation-based properties. Correlating label-free representation-quality metrics with layer-wise performance across 15 downstream tasks, we find that several metrics track layer quality for genre classification, emotion recognition, automatic tagging, and beat tracking, albeit with varying strength across tasks and pre-training paradigms. However, all metrics fail on tonal tasks such as key estimation and chord recognition, indicating that no single property serves as a general proxy for representation quality across music information retrieval tasks. To address this gap, we introduce a pitch-transposition equivariance measure that captures properties missed by these standard metrics, providing a consistent indicator of tonal quality across model families. Finally, we show that intrinsic metrics can serve as effective proxies for layer selection, matching or outperforming trainable multi-layer fusion methods, particularly in limited-data settings.
Primary: Universitat Pompeu Fabra (Music Technology Group)
All Institutions: Universitat Pompeu Fabra, Music Technology Group
This paper makes a significant contribution to the understanding of representation quality in music foundation models by introducing a systematic layer-wise analysis and a novel pitch-transposition equivariance metric, demonstrating that intrinsic properties can effectively guide layer selection, particularly for tonal tasks where standard metrics fail.
The paper presents a rigorous and systematic layer-wise analysis of 12 music foundation models across three distinct pre-training paradigms (masked, autoregressive, contrastive). The methodology is sound, employing a comprehensive suite of label-free intrinsic metrics (Intrinsic Dimension, Curvature, Anisotropy, Effective Rank, LiDAR, InfoNCE) to characterize representation geometry. The key methodological contribution is the introduction of Pitch-Transposition Equivariance (PTE), a novel metric designed to capture tonal structure that standard geometric metrics miss. The approach of correlating these intrinsic properties with downstream probe performance across 15 diverse MIR tasks provides a robust framework for understanding representation quality without relying solely on task-specific supervision.
The experimental evaluation is extensive, covering a wide range of downstream tasks including tonal, rhythmic, timbral, semantic, and similarity tasks. The results are well-supported, demonstrating that while standard metrics correlate well with non-tonal tasks, they fail for tonal tasks, necessitating the new PTE metric. The paper effectively demonstrates that intrinsic metrics can serve as effective proxies for layer selection, often outperforming trainable multi-layer fusion methods, particularly in low-data regimes. The analysis of depth-wise trends across different model families provides valuable insights into how representation properties evolve.
The paper provides significant detail on the models, datasets (MTG-Jamendo, GiantSteps, NSynth, etc.), and evaluation protocols. The code and extended results are available on the project page, enhancing reproducibility. The use of standard datasets and publicly available models facilitates independent verification.
The study is correlational; it identifies properties associated with good performance but does not establish causality. The metrics are evaluated on frozen representations, so their utility for guiding pre-training or fine-tuning is not directly addressed. The analysis is limited to 12 models, which, while diverse, may not cover all architectural variations. The PTE metric, while promising, is specific to tonal tasks and may not generalize to other musical attributes.
This work provides practical guidelines for selecting layers in music foundation models, potentially reducing the computational cost of model evaluation and deployment. It advances the theoretical understanding of representation quality in audio models, bridging the gap between geometric analysis and practical MIR performance. The findings are relevant to the broader community working with self-supervised audio representations. This paper makes a significant contribution to the understanding of representation quality in music foundation models by introducing a systematic layer-wise analysis and a novel pitch-transposition equivariance metric, demonstrating that intrinsic properties can effectively guide layer selection, particularly for tonal tasks where standard metrics fail.
Spoken dialogue is a natural form of human--computer interaction, yet most speech language models remain limited to turn-based operation and lack real-time adaptability, such as user barge-in. Recent duplex speech-to-speech and speech-to-text models reduce latency by replacing multi-stage pipelines, but often compromise speech quality because accurate ASR, interruption handling, and high-fidelity synthesis must be optimized jointly. We propose VoiceChat-TTS, a low-latency, continuous, and streamable text-to-speech model for interactive agents. VoiceChat-TTS is driven directly by LLM text-token streams, supports explicit interruption via control tokens, and produces silence when no textual input is available. The model enables always-on, responsive speech generation while preserving modularity and high speech quality, and it supports mid-utterance interruptions without resetting the KV cache.
Primary: NVIDIA Corporation
All Institutions: NVIDIA Corporation
VoiceChat-TTS presents a robust, low-latency, and interruptible streaming TTS system that effectively bridges the gap between high-quality offline synthesis and the real-time demands of full-duplex interactive agents. The comprehensive evaluation, including detailed latency analysis and interruption handling benchmarks, demonstrates its practical utility and sets a strong baseline for future research in continuous speech synthesis.
The paper proposes VoiceChat-TTS, a continuous, streamable TTS model built upon the Audio Flamingo 3-Chat architecture. The core technical contributions involve adapting a streaming decoder for full-duplex interaction. Key modifications include: 1) A Character-Aware Subword Encoder to handle out-of-vocabulary subwords from LLMs by converting them to character sequences processed by a shallow Transformer. 2) An interruption mechanism using explicit control tokens to halt generation and transition to silence without resetting the KV cache. 3) Audio Prompt Conditioning using a 3-second reference audio to stabilize speaker identity at the start of generation, addressing the lack of context in streaming settings. 4) A Mixture of Gaussian Estimation Head (MoGH) for faster RVQ token decoding. The approach is pragmatic, focusing on engineering solutions (gated fusion, specific tokenization, silence modeling) to enable real-time, interruptible speech synthesis. While not fundamentally novel in terms of architecture (it extends existing streaming decoders), the specific integration of these components for a unified duplex TTS system is a valuable engineering contribution.
The evaluation is comprehensive and well-structured. The authors compare VoiceChat-TTS against strong offline (Chatterbox-TTS, Qwen3-TTS) and streaming (Audio Flamingo 3-Chat) baselines. Metrics include CER, WER, SECS (speaker similarity), and Squim-MOS. The results show that VoiceChat-TTS achieves competitive quality, significantly outperforming the base Audio Flamingo 3-Chat decoder in intelligibility and quality. The paper also includes a dedicated interruption evaluation using the Full-Duplex-Bench (FDB) subset, measuring Stop Latency and Leakage. The latency analysis is particularly strong, providing detailed breakdowns of acoustic-token ITL and codec decoding time on specific hardware (RTX A6000), demonstrating a 2.1x speedup over a comparable baseline. The use of both synthetic and real conversational data for training is well-justified.
The paper provides significant detail for reproduction. It specifies the model size (977M parameters), the base architecture (Gemma 3-based), the codec configuration (12.5 Hz, 31-codebook RVQ), and the training stages (pretraining on single-turn data, fine-tuning on multi-turn). The code is publicly available via NVIDIA NeMo Speech, and the model checkpoint is on Hugging Face. The training data sources are listed (LibriTTS, HiFiTTS, synthetic data, Fisher). The evaluation protocol is also clearly defined, including the use of Silero VAD for interruption timing detection. This high level of detail ensures that the work is reproducible.
The authors acknowledge several limitations. First, the model lacks user-audio conditioning, meaning it cannot dynamically adapt its prosody to the user's speech characteristics or interruptions in real-time (it relies on text tokens). Second, there is a lack of controlled component-wise ablations; while preliminary experiments suggested cumulative gains, a systematic ablation study is missing. Third, speaker similarity (SECS) degrades over longer sequences of continuous generation, particularly for unseen speakers, indicating a challenge in maintaining long-context speaker consistency. Finally, the evaluation of "silence" generation relies on ASR-based metrics, which may not fully capture the perceptual quality of silence or the naturalness of the transition into/out of silence.
VoiceChat-TTS contributes to the development of more natural and responsive human-computer interaction systems. By enabling low-latency, interruptible speech synthesis, it facilitates the deployment of always-on voice assistants and conversational agents that behave more like human interlocutors. This has positive implications for accessibility, customer service, and companion AI. However, the ease of generating high-quality, realistic speech also raises concerns about potential misuse in deepfakes or deceptive interactions, although the model's reliance on text input mitigates some of the risks associated with direct audio-to-audio generation. VoiceChat-TTS presents a robust, low-latency, and interruptible streaming TTS system that effectively bridges the gap between high-quality offline synthesis and the real-time demands of full-duplex interactive agents. The comprehensive evaluation, including detailed latency analysis and interruption handling benchmarks, demonstrates its practical utility and sets a strong baseline for future research in continuous speech synthesis.
Vocalized audio synthesis, the task of generating audio in which intelligible speech is embedded within an environmental soundscape, underpins applications such as podcast production and video dubbing. Existing Text-to-Audio (T2A) systems either reduce quoted speech to unintelligible vocal murmur or delegate it to a separate TTS model with post-hoc mixing, which forfeits control over when speech occurs and how it interacts with the scene. We present VoxAudio, a causal autoregressive flow matching model that addresses this problem from three complementary aspects. At the architecture level, chunk-wise causal factorization with independent per-chunk noise levels lets audio be emitted through sliding-window streaming inference with KV caching at variable target durations; to enable inference at arbitrary chunk granularities, we further pretrain the model with randomized chunk boundaries. At the preference level, multi-reward Negative-aware FineTuning (NFT) jointly optimizes semantic fidelity, linguistic accuracy, aesthetic quality, and temporal grounding At the data level, to supply the missing supervision for vocal content, we build VoxCorpus, a large-scale corpus whose captions quote the verbatim transcript of embedded speech with time intervals, and VoxBench, an interval-annotated benchmark with a temporal-grounding metric. Experiments on four benchmarks spanning general audio, speech, and unified vocalized audio validate the effectiveness and efficiency of VoxAudio. Our code and demos are available at https://voxaudio.github.io.
Primary: Zhejiang University
All Institutions: Zhejiang University
The paper presents a technically sound approach to streaming audio generation with speech, introducing valuable data and benchmarks. The adaptation of autoregressive flow matching with chunk-wise causal factorization is a meaningful contribution to efficient audio synthesis, and the multi-reward alignment strategy offers a robust path for preference optimization in continuous domains.
The paper proposes VoxAudio, a causal autoregressive flow matching model for vocalized audio synthesis. The core technical contribution lies in adapting flow matching to a streaming, autoregressive setting. Specifically, it employs chunk-wise causal factorization with independent per-chunk noise levels, allowing for sliding-window inference with KV caching. This addresses the train-test gap often seen in causal diffusion/flow models by randomizing chunk boundaries during pretraining. The methodology also includes a multi-reward Negative-aware FineTuning (NFT) stage to align the model with human preferences across semantic, linguistic, aesthetic, and temporal grounding dimensions. The architectural adaptations (causal convolutions, chunk-wise causal attention) are sound and necessary for the stated streaming objective. The integration of NFT for audio generation is a notable extension of recent RLHF techniques to continuous flow-matching domains.
The authors introduce VoxCorpus, a new dataset with verbatim speech annotations and time intervals, and VoxBench, a benchmark with a temporal-grounding metric. Experiments are conducted on four benchmarks. The results claim superiority in speech fidelity and timing precision over baselines. However, the provided text is truncated, preventing a full assessment of the quantitative results (e.g., FAD, UTMOS, WER scores) and ablation studies. The claim of "outperforming current vocalized audio generation baselines" is significant but requires rigorous verification against strong baselines like separate TTS+T2A pipelines or unified models like Dasheng-AudioGen. The introduction of a new benchmark is a strong positive for the community, provided the annotation quality and metric validity are high.
The paper provides a code and demo link. The methodology describes the use of pre-trained components (Universe Audio VAE, T5, Whisper, CLAP), which aids reproducibility. The specific details of the NFT reward weights and the exact implementation of the randomized chunk boundary training are crucial for reproduction and are likely detailed in the full text (which is partially truncated here). The use of standard open-source models for evaluation (Whisper, CLAP) ensures that the evaluation protocol is replicable.
The paper does not explicitly discuss the computational cost of the autoregressive flow matching compared to non-causal diffusion models, which is a significant factor for practical deployment. The reliance on multiple external models (Whisper for WER, CLAP for semantic reward) for NFT introduces potential biases and error propagation. The quality of the VoxCorpus dataset, particularly the accuracy of the "verbatim transcript" and time intervals, is critical and any errors there would limit the model's potential. The truncation of the text prevents assessment of failure cases or specific limitations in handling complex acoustic scenes.
VoxAudio addresses a significant gap in audio generation by enabling intelligible speech within environmental soundscapes, which has applications in podcasting, video dubbing, and immersive media. The release of VoxCorpus and VoxBench provides valuable resources for the community. However, the ability to generate realistic speech in arbitrary contexts raises concerns about misuse for deepfakes or misinformation, although the environmental context may mitigate some immediate risks compared to pure TTS. The paper presents a technically sound approach to streaming audio generation with speech, introducing valuable data and benchmarks. The adaptation of autoregressive flow matching with chunk-wise causal factorization is a meaningful contribution to efficient audio synthesis, and the multi-reward alignment strategy offers a robust path for preference optimization in continuous domains.
Modern text-to-speech (TTS) is dominated by autoregressive (AR) codec language models, whose left-to-right decoding brings latency that grows with utterance length, error accumulation along the committed prefix, and an artificial generation order imposed on the Residual Vector Quantization (RVQ) token grid. We propose Luna-TTS Family, diffusion-language-model-based TTS systems pretrained on 1 million hours of speech across Chinese, English, Japanese, and Korean. The family is built by progressive adaptation of a pretrained AR text LLM, from causal to bidirectional and finally to block-causal attention, and comprises two variants sharing a single tokenizer, data pipeline, and 0.6B backbone lineage. Luna-TTS is fully non-autoregressive: it generates the entire RVQ token grid in a fixed number of parallel refinement steps, with zero-shot voice cloning and speech editing arising natively as infilling. Luna-TTS Realtime, derived by continual training, is autoregressive over blocks of 32 codec frames (1.28s) while denoising each block in parallel; it supports KV-cached blockwise generation and incremental audio delivery, achieving an end-to-end RTF of 0.0240 and 41.6 ms local first-block latency under the warmed serving protocol. An annealed fine-tuning stage adds explicit control over emotion and non-verbal vocalizations (NVVs), and a reinforcement-learning stage applies GRPO with policy ratios computed over the realized denoising trajectory. On Seed-TTS-Eval, Luna-TTS achieves the best results on all four metrics among compared open-source and commercial systems (0.73 CER / 79.7 SIM on test-zh, 1.49 WER / 76.8 SIM on test-en); on the harder in-the-wild CV3-Eval, it posts the lowest Mandarin and English error rates in our comparison. Against leading commercial systems, it achieves the best results on most objective, model-based, and human-rated metrics for NVV and emotion control.
Primary: VUI Labs Research
All Institutions: VUI Labs Research
Luna-TTS introduces a scalable, block-diffusion-based TTS framework that achieves state-of-the-art quality and latency by adapting pretrained LLMs through progressive architectural changes and RL post-training, effectively resolving the latency-quality trade-off inherent in previous diffusion TTS systems.
The paper proposes a novel architecture for Text-to-Speech (TTS) by adapting Large Language Models (LLMs) into Diffusion Language Models (dLLMs). Specifically, it introduces a "block-causal" attention mechanism that allows for streaming generation while retaining the parallel denoising benefits of diffusion models. The approach involves a progressive adaptation from causal to bidirectional and finally to block-causal attention on a pretrained AR text LLM. A key technical contribution is the application of Group Relative Policy Optimization (GRPO) directly over the realized denoising trajectory, optimizing for content correctness and speaker similarity. The use of a semantic-distilled RVQ tokenizer where the first codebook is anchored to linguistic content is a significant methodological detail that facilitates the diffusion process. The transition from fully parallel masked diffusion to block-autoregressive diffusion for latency reduction is a well-reasoned engineering solution to a known bottleneck in diffusion-based generation.
The evaluation is comprehensive, covering both objective metrics (CER, WER, SIM) and subjective/human-rated metrics for emotion and non-verbal vocalization (NVV) control. The paper claims state-of-the-art performance on Seed-TTS-Eval and CV3-Eval, outperforming both open-source and leading commercial systems. The inclusion of a "warmed serving protocol" to measure real-time factor (RTF) and first-block latency adds practical relevance to the experimental setup. The comparison against commercial baselines strengthens the claim of production-readiness. However, the specific numerical results are redacted in the provided text, preventing a precise verification of the magnitude of improvement, though the qualitative claims are strong.
The paper provides detailed descriptions of the architecture, including the tokenizer design, the masked diffusion formulation, and the training schedule. The mention of "0.6B backbone lineage" and "1 million hours of speech" provides scale context. However, the specific hyperparameters for the GRPO stage and the exact implementation details of the block-causal attention masking are likely critical for reproduction and may be subject to interpretation. The lack of publicly released code or weights (implied by "none" for URLs) significantly hinders immediate reproducibility for the broader community, although the technical report format suggests a high level of detail.
The primary limitation is the reliance on a specific RVQ tokenizer and the assumption that the semantic anchoring of the first codebook generalizes well across all languages and domains. The block-diffusion approach, while reducing latency, still requires multiple denoising steps per block, which may not be as fast as single-step AR decoding in extreme low-latency scenarios. Furthermore, the complexity of training a diffusion LLM with RL post-training is significantly higher than standard AR TTS training, potentially limiting accessibility. The paper does not extensively discuss the failure modes of the NVV control or the robustness of the emotion conditioning under adversarial text inputs.
This work represents a significant step towards making diffusion-based generative models viable for real-time, production-grade audio applications. By bridging the gap between the quality of diffusion models and the latency requirements of streaming TTS, it expands the toolkit available for developers building voice assistants, gaming NPCs, and accessibility tools. The emphasis on expressive control (emotion, NVVs) aligns with the growing demand for more natural and engaging human-computer interaction. However, the potential for misuse in deepfake generation remains a concern, necessitating robust safety measures in deployment. Luna-TTS introduces a scalable, block-diffusion-based TTS framework that achieves state-of-the-art quality and latency by adapting pretrained LLMs through progressive architectural changes and RL post-training, effectively resolving the latency-quality trade-off inherent in previous diffusion TTS systems.
Long-form song generation models continue to improve in duration, structural integrity, and acoustic complexity, making reliable aesthetic rewards increasingly important for aligning these models with human preferences. However, reward models for complete songs remain limited, and existing evaluators typically predict scores in a single forward pass without providing readable explanations. We introduce MUSECRITIC, a semi-scalar reward model that generates a natural-language critique covering five aesthetic dimensions and uses it as an intermediate representation to predict continuous reward scores. MUSECRITIC follows a two-stage training pipeline: a teacher model first provides high-quality critiques for supervised fine-tuning, after which the fine-tuned model generates its own critiques for reward learning, mitigating distribution shift between training and inference. On an in-domain test set of 200 SongEval songs, MUSECRITIC reduces macro-averaged mean squared error from 0.2875 to 0.2316 and improves macro-averaged LCC, SRCC, and Kendall's tau to 0.9068, 0.8838, and 0.7178, respectively. On the out-of-domain Music Arena benchmark with 733 preference pairs, it achieves the highest accuracy of 71.35%. Moreover, using MUSECRITIC with GRPO improves Muse-0.6B on all nine aesthetic metrics from SongEval and Audiobox Aesthetics. These results demonstrate that critique-conditioned reward modeling reduces scoring error and provides an effective optimization signal for song generation. The project repository is available at https://github.com/WuqnEl/MuseCritic.
Primary: Fudan University
All Institutions: Fudan University
MuseCritic introduces a critique-conditioned reward modeling framework for long-form song generation, demonstrating that natural-language aesthetic critiques serve as effective intermediate representations for improving score prediction accuracy and enabling successful reinforcement learning alignment.
The paper proposes MuseCritic, a semi-scalar reward model for long-form song generation that integrates natural-language aesthetic critiques as an intermediate representation before predicting continuous scores. The methodology involves a two-stage training pipeline: first, supervised fine-tuning (SFT) of an audio-language model backbone using critiques generated by a powerful external teacher (Gemini-3-Pro) conditioned on expert scores; second, reward learning where the model generates its own critiques (self-generated) to mitigate distribution shift, followed by training a reward head to predict scores conditioned on these self-generated critiques. This "critique-then-score" approach is conceptually sound and aligns with recent trends in LLM-as-a-Judge and reasoning-enhanced reward modeling. However, the novelty is somewhat limited by the reliance on a black-box teacher for data generation and the use of standard LoRA fine-tuning on a large foundation model, rather than proposing a fundamentally new architectural primitive for audio processing.
The experimental evaluation is comprehensive and rigorous. The authors evaluate on in-domain SongEval metrics (MSE, LCC, SRCC, KTAU) and out-of-domain Music Arena preference accuracy. MuseCritic consistently outperforms baselines, including a retrained SongEval (UTMOS) baseline and Gemini-3.1-Pro. Crucially, the ablation studies effectively isolate the contributions of the critique generation, self-generated critiques vs. offline critiques, and SFT initialization. The downstream reinforcement learning experiment using GRPO to optimize Muse-0.6B with MuseCritic as the reward model demonstrates practical utility, showing improvements across multiple aesthetic metrics. The use of deterministic decoding and fixed-prefix evaluation windows adds robustness to the downstream results.
The paper provides detailed implementation details, including hyperparameters, training configurations, and data splitting strategies. The code repository is linked. The use of a specific external teacher (Gemini-3-Pro) for data generation is a potential reproducibility hurdle for others without access to that specific model version, but the prompt templates are provided. The data splitting is clearly defined.
The primary limitation is the computational cost and latency introduced by the autoregressive critique generation step, which makes inference slower than direct score regression. Additionally, the method relies heavily on the quality of the external teacher for initial data generation, which may introduce biases or hallucinations if not properly verified (though the authors claim human verification). The evaluation is somewhat limited to Chinese and English vocal songs, and the generalizability to other genres or languages is not fully explored.
This work contributes to the field of AI-generated content evaluation and alignment. By providing a more interpretable and potentially more accurate reward model for song generation, it facilitates better optimization of generative models via reinforcement learning. This can lead to higher quality AI-generated music and more reliable evaluation benchmarks. The emphasis on interpretable critiques also aligns with the broader goal of making AI systems more transparent and trustworthy. MuseCritic introduces a critique-conditioned reward modeling framework for long-form song generation, demonstrating that natural-language aesthetic critiques serve as effective intermediate representations for improving score prediction accuracy and enabling successful reinforcement learning alignment.
Prompted attribute agreement is widely used as evidence of text-to-music controllability, yet a requested attribute may occur simply because it is already common in the model's output distribution. We introduce a matched counterfactual evaluation that separates target occurrence from instruction-attributable control. Each family contains a neutral input that omits the scored attribute and two otherwise matched inputs that swap the requested target. All three are rendered through frozen native-interface adapters with a shared seed. Applied to global key and beat grouping in three open systems, this design changes the empirical conclusion. ACE-Step 1.5 and Stable Audio 3 Medium exhibit substantial key control, whereas LeVo2 does not. For beat grouping, the same models redirect toward the rare three-beat target, but high four-beat agreement is largely inherited from neutral outputs: Stable Audio 3 produces four-beat grouping in 0.97 of neutral cases but only 0.56 under its explicit four-beat treatment. Off-attribute placebos, external recognizer validation, blind expert annotation, and multi-seed sentinels support the attribution. When targets have unequal output priors, agreement describes what a model produced, while matched neutral and target-swap contrasts test whether the instruction changed it.
Primary: Unknown
All Institutions: Unknown
The paper presents a novel counterfactual evaluation framework that rigorously distinguishes between attribute occurrence and instruction-attributable control in text-to-music models, revealing that high agreement rates often reflect data priors rather than genuine controllability, thereby providing a more accurate and actionable benchmark for the field.
The paper introduces a rigorous counterfactual evaluation framework for text-to-music models, specifically targeting the distinction between "prompted attribute agreement" (mere occurrence) and "instruction-attributable control" (causal influence of the prompt). The core methodological contribution is the "matched neutral--A--B contrast family" design. By generating audio with a neutral prompt (omitting the target attribute) and two treatment prompts (specifying different target values) using a shared seed and frozen adapters, the authors isolate the effect of the instruction from the model's inherent output priors. This approach is statistically sound and addresses a critical flaw in current evaluation practices where high agreement rates may simply reflect the model's bias toward common attributes (e.g., 4/4 time). The metrics proposed—Treatment Agreement, Enhancement ($\Delta$), and Margin—are well-defined and provide a granular view of controllability.
The evaluation is applied to three prominent open systems: ACE-Step 1.5, Stable Audio 3 Medium, and LeVo2. The experiments cover global key and beat grouping. The results are compelling and counter-intuitive to standard benchmarks: while ACE-Step and Stable Audio 3 show strong key control, LeVo2 shows little attributable response. More importantly, for beat grouping, the paper demonstrates that high four-beat agreement in Stable Audio 3 is largely inherited from neutral outputs (0.97 neutral rate), and the explicit instruction actually *decreases* the probability of four-beat grouping (negative $\Delta$). This finding fundamentally changes the empirical conclusion about these models' capabilities. The use of off-attribute placebos, external recognizer validation, blind expert annotation, and multi-seed sentinels adds significant robustness to the claims.
The paper provides extensive details on the experimental setup, including frozen native-interface adapters, specific decoding settings, and seed handling. The authors release a code-and-data package with instantiated cases, prompts, configurations, and analysis code. The use of deterministic seeds and frozen model revisions ensures that the generation process is reproducible. The detailed appendix provides literal renderings of inputs and full provenance, making it highly reproducible.
The evaluation is limited to global key and beat grouping. It does not address local structural edits, modulation, or continuous controls. The neutral contrast necessarily changes the rendered carrier (e.g., omitting a field vs. filling it), which introduces a confound between the absence of the attribute and the change in prompt structure. The authors acknowledge this and use A/B swaps and off-attribute placebos to mitigate it, but a factorial design crossing carrier templates would be stronger. The evaluation is also limited to the specific public interfaces of the models; internal architectural changes that might improve control are not explored.
This work has significant implications for the development and evaluation of generative audio models. By establishing that "agreement" is not sufficient evidence of "control," it provides a new standard for benchmarking. This can guide model developers to focus on improving genuine instruction-following capabilities rather than just aligning with data priors. It also helps practitioners choose models based on their actual controllability profiles (e.g., good at rare targets vs. common targets). The framework is generalizable to other multimodal generation tasks where output priors are strong. The paper presents a novel counterfactual evaluation framework that rigorously distinguishes between attribute occurrence and instruction-attributable control in text-to-music models, revealing that high agreement rates often reflect data priors rather than genuine controllability, thereby providing a more accurate and actionable benchmark for the field.
Generating coherent audio scenes that simultaneously blend speech, music, and sound effects remains a significant challenge. Current approaches typically rely on a disjointed pipeline where a frozen, decoupled text encoder feeds a separate audio decoder, limiting cross-modal optimization and leading to poor speech intelligibility. To overcome these limitations, we introduce MiDashengLM-Gen, an end-to-end framework that couples a pre-trained Large Language Model (LLM) with per-token conditional flow matching for autoregressive, variable-length mixed-audio scene generation. MiDashengLM-Gen represents a first approach for general text-to-audio generation with one end-to-end trained model. Empirical evaluations demonstrate that MiDashengLM-Gen drastically improves speech intelligibility over existing unified models. On the Seed-TTS benchmark, English Word Error Rate (WER) drops from 12.15% to 2.79%, approaching the performance of dedicated Text-to-Speech (TTS) systems (1.24%). Furthermore, the framework extends effectively to multilingual settings, yielding highly competitive multilingual WERs compared to existing baselines. Lastly, the model maintains competitive mixed-audio generation quality on the MECAT benchmark. Code and checkpoints are available at https://github.com/xiaomi-research/midashenglm-gen and https://huggingface.co/mispeech/midashenglm-gen, and the demo page is available at https://xingws.github.io/midashenglm-gen-demo/.
Primary: Xiaomi Inc.
All Institutions: MiLM Plus, Xiaomi Inc., X-LANCE Lab, Shanghai Jiao Tong University
[One sentence main contribution]. MiDashengLM-Gen introduces an end-to-end LLM-based autoregressive framework with per-token flow matching for unified mixed-audio scene generation, achieving state-of-the-art speech intelligibility and competitive mixed-audio quality. [Comprehensive analysis of the technical contribution, methodology, and significance to the field]. The paper presents a significant step forward in unified audio generation by effectively integrating large language models with continuous flow matching. The key insight regarding the DiT width constraint provides valuable theoretical and practical guidance for training such hybrid architectures. The substantial improvement in speech intelligibility addresses a major bottleneck in previous unified models, making the generated audio more usable for practical applications. While the approach builds on existing components (LLMs, Flow Matching, Tokenizers), the specific integration for variable-length mixed-audio generation is novel and well-validated.
The paper proposes MiDashengLM-Gen, an end-to-end framework for unified audio scene generation (speech, music, sound effects). The core architectural innovation is replacing the frozen text encoder and non-autoregressive diffusion backbone of its predecessor (Dasheng AudioGen) with a pre-trained Large Language Model (Qwen3-1.7B) coupled with per-token conditional flow matching. The method utilizes a structured multi-view captioning approach to decompose audio scenes into semantic views (global, transcript, SFX, music, env, speaker). A key technical contribution is the identification of a convergence prerequisite: the DiT decoder width must strictly exceed the audio latent dimensionality. The approach integrates an audio-text alignment stage to map high-dimensional audio latents into the LLM's token space. While the combination of LLMs and flow matching is not entirely new, applying it to *unified mixed-audio scene generation* with *per-token* flow matching for variable-length output is a distinct methodological shift from previous fixed-length or disjointed pipeline approaches.
The evaluation is comprehensive, covering single-type (AudioCaps, MusicCaps) and mixed-type (MECAT) generation, as well as speech intelligibility (Seed-TTS, multilingual WER/CER). The results show significant improvements in speech intelligibility (WER drop from 12.15% to 2.79% on Seed-TTS) compared to the previous Dasheng AudioGen model. The model maintains competitive performance on mixed-audio metrics (FAD, FD, KL) on MECAT, although it trails dedicated sound-effect models like TangoFlux on pure sound effects, which is acknowledged. The ablation studies effectively validate the necessity of the audio-text alignment stage and the DiT width constraint. The use of standard benchmarks (Seed-TTS, MECAT) ensures fair comparison.
The paper provides detailed implementation details, including model sizes (Qwen3-1.7B, DiT 16 layers), training hyperparameters (batch size, LR, epochs), and data sources. Code and checkpoints are made publicly available via GitHub and Hugging Face, significantly enhancing reproducibility. The description of the structured captioning and the specific flow matching objective is sufficiently detailed for replication.
The authors acknowledge several limitations: variable-length generation is bounded by training data distribution (1-20 seconds); speech intelligibility still trails dedicated TTS systems (especially in low-resource languages); and the model supports only coarse speaker-style control without explicit voice cloning or fine-grained temporal control. The trade-off between mixed-audio coordination and single-source acoustic fidelity (e.g., trailing TangoFlux on sound effects) is also a noted limitation.
This work advances the field of generative audio by demonstrating that LLM-based autoregressive frameworks can effectively handle complex, mixed-audio scene generation with high speech intelligibility. This has significant implications for immersive media, gaming, and film production where coherent multi-source audio is required. The open-source release contributes to the community's ability to build upon unified audio generation models. [One sentence main contribution]. MiDashengLM-Gen introduces an end-to-end LLM-based autoregressive framework with per-token flow matching for unified mixed-audio scene generation, achieving state-of-the-art speech intelligibility and competitive mixed-audio quality. [Comprehensive analysis of the technical contribution, methodology, and significance to the field]. The paper presents a significant step forward in unified audio generation by effectively integrating large language models with continuous flow matching. The key insight regarding the DiT width constraint provides valuable theoretical and practical guidance for training such hybrid architectures. The substantial improvement in speech intelligibility addresses a major bottleneck in previous unified models, making the generated audio more usable for practical applications. While the approach builds on existing components (LLMs, Flow Matching, Tokenizers), the specific integration for variable-length mixed-audio generation is novel and well-validated.
We present RT-SEMamba, a fully causal speech enhancement (SE) model built upon causal time-frequency Mamba blocks. Unlike Transformer-based architectures that rely on a growing key-value cache, Mamba propagates a fixed-size recurrent state per layer, enabling memory- and bandwidth-efficient long-form inference. We further introduce a progressive knowledge distillation (KD) strategy that compresses an 8-layer teacher into a shallow 1-layer student by jointly distilling complex spectral outputs and intermediate representations. On Voicebank-DEMAND, the 8-layer RT-SEMamba achieves 3.32 PESQ with a 25 ms algorithmic latency constraint, and the distilled 1-layer student improves over a naive 1-layer baseline from 3.06 to 3.18 PESQ while preserving the same steady-state RTF, delivering a 2.75x speedup over the teacher. These results demonstrate that state-space models with progressive KD provide a competitive quality-latency trade-off for real-time SE.
Primary: National Taiwan University
All Institutions: Academia Sinica, National Taiwan University, Kore University of Enna, University of Palermo, NVIDIA
The paper presents RT-SEMamba, a causal Mamba-based speech enhancement model with progressive knowledge distillation, offering a competitive quality-latency trade-off for real-time applications.
The paper proposes RT-SEMamba, a fully causal speech enhancement model utilizing Time-Frequency Mamba (TF-Mamba) blocks. The core methodological contribution lies in adapting the selective state-space model (SSM) architecture for strict real-time streaming, specifically by ensuring causality in the temporal dimension while maintaining bidirectional frequency modeling. The authors introduce a progressive knowledge distillation (KD) strategy to compress an 8-layer teacher into a 1-layer student. This involves distilling both complex spectral outputs (magnitude, phase, complex spectrum) and intermediate feature representations. The methodology is technically sound and addresses a relevant gap in applying Mamba architectures to streaming audio, where memory efficiency is critical. However, the novelty is somewhat incremental; adapting existing SSM blocks for causality and applying standard KD techniques are well-established practices in the field, though their specific combination and tuning for SE are valuable.
The experiments are conducted on the VoiceBank-DEMAND dataset, a standard but relatively small benchmark for SE. The results show that the distilled 1-layer student achieves 3.18 PESQ, outperforming a naive 1-layer baseline (3.06) and approaching the 8-layer teacher (3.32). The paper provides a detailed analysis of the quality-latency trade-off, demonstrating that KD effectively shifts the Pareto frontier. The inclusion of hybrid Mamba-Transformer ablations adds depth to the architectural analysis. However, the evaluation is limited to objective metrics (PESQ, CSIG, etc.) on a single dataset. The lack of subjective listening tests or evaluation on larger, more diverse benchmarks (like DNS Challenge datasets) limits the generalizability of the claims. The comparison with prior works is fair, but the performance gains, while consistent, are modest in absolute terms.
The paper provides sufficient architectural details, including layer configurations, loss functions, and training hyperparameters (e.g., ramp-up steps). The code is promised to be released on GitHub, which enhances reproducibility. The dataset and evaluation protocol are standard, facilitating independent verification. The description of the streaming inference setup, including buffer management, is clear enough for implementation.
The primary limitation is the reliance on the VoiceBank-DEMAND dataset, which may not reflect performance on more challenging, real-world noise conditions or larger-scale training data. The model's performance on unseen, complex acoustic environments is not thoroughly tested. Additionally, the "progressive" nature of the KD is described but the specific scheduling or adaptive mechanisms are standard, limiting the perceived innovation in the distillation strategy itself. The paper does not discuss potential failure modes of the Mamba blocks in handling very long-term dependencies compared to Transformers in the streaming context.
This work contributes to the development of efficient, real-time audio processing systems, which are crucial for accessibility technologies (hearing aids), immersive communications (AR/VR), and edge computing applications. By demonstrating that Mamba-based models can achieve competitive performance with lower latency and memory footprint than Transformers, it encourages the exploration of alternative sequence modeling architectures in audio AI. The open-source release will further facilitate research in this direction. The paper presents RT-SEMamba, a causal Mamba-based speech enhancement model with progressive knowledge distillation, offering a competitive quality-latency trade-off for real-time applications.