Human face-to-face conversation requires coordinating speech, gestures and expressions across both speaking and listening. We present ELLSA 2, an end-to-end full-duplex dialogue model that jointly optimizes speech interaction and streaming gesture-expression generation. An autoregressive, chunk-wise synthesizer combines pretrained acoustic features with distinct listening and speaking hidden states from the dialogue backbone and speech synthesizer, enabling continuous non-verbal generation even when the model is verbally silent. To translate dialogue understanding into vivid semantic gestures, we introduce contextual semantic calls and template guidance, separating whether to invoke an available semantic gesture from how to realize it. We also establish a curated data and evaluation pipeline for dialog-aware co-speech gesture generation and investigate reference-based multimodal large language models (MLLMs) as automatic judges. Experiments show that ELLSA 2 remains highly competitive with open-source full-duplex speech interaction models and outperforms cascaded baselines across dialog-aware gesture and expression metrics. Semantic template guidance further significantly improves semantic gesture quality. Strong system-level correlations with human judgments support these reference-based scores as useful preliminary metrics for full-duplex and semantic gesture assessment. Together, these results advance dialogue modeling from spoken responses toward coordinated multimodal interaction.
Primary: Tsinghua University
All Institutions: Tsinghua University, ByteDance
ELLSA 2 introduces a unified end-to-end framework for full-duplex speech interaction with coordinated gesture and expression generation, featuring a novel semantic template guidance mechanism that significantly improves the quality and appropriateness of non-verbal behaviors in dialogue.
The paper proposes ELLSA 2, a unified end-to-end framework that integrates full-duplex speech interaction with streaming gesture and facial expression generation. The core methodological contribution is the "contextual semantic call" mechanism, which decouples the decision of *whether* to perform a semantic gesture from *how* to realize it. This is achieved via a binary classification head on the LLM backbone's hidden states, which triggers the retrieval of a curated exemplar template to guide the autoregressive gesture generator. The architecture uses a chunk-wise autoregressive synthesizer that interleaves speech, listening, and motion tokens, allowing non-verbal behaviors to persist during verbal silence (listening states). The use of RVQ-VAE for motion tokenization and a curriculum training strategy to integrate modalities is technically sound and well-motivated.
The experimental setup is robust, introducing a new human-verified benchmark (138 clips) specifically for dialog-aware gesture generation, covering full-duplex behaviors and semantic gestures. The evaluation employs both objective metrics (prosodic alignment, activity ratio) and subjective human evaluation, along with MLLM-as-a-judge metrics. The results demonstrate that ELLSA 2 outperforms cascaded baselines and remains competitive with state-of-the-art full-duplex speech models. The semantic template guidance shows a significant improvement in semantic gesture quality (human score 2.014 to 2.711). The correlation analysis between MLLM judges and human annotators (0.853/0.919) provides strong evidence for the utility of the proposed automatic metrics.
The paper provides extensive implementation details, including tokenizer specifications, training hyperparameters, and data curation pipelines. The authors commit to releasing code, data, and model checkpoints upon acceptance. The detailed description of the semantic exemplar library construction and the evaluation protocols enhances reproducibility.
The system relies on a curated library of 3,227 exemplars for semantic gestures, which may limit the diversity of semantic actions compared to fully generative approaches. The evaluation, while improved, is still based on a relatively small test set (138 clips). The model's performance on long-horizon coherence of gestures over extended conversations is not deeply analyzed.
This work advances the field of human-computer interaction by moving beyond text or speech-only interfaces to fully embodied, multimodal agents. The proposed evaluation metrics and benchmark will be valuable for the community working on co-speech gesture generation and full-duplex dialogue systems. The integration of semantic understanding with physical gesture generation has implications for robotics, virtual avatars, and immersive communication. ELLSA 2 introduces a unified end-to-end framework for full-duplex speech interaction with coordinated gesture and expression generation, featuring a novel semantic template guidance mechanism that significantly improves the quality and appropriateness of non-verbal behaviors in dialogue.
Target speaker extraction (TSE) isolates a desired speaker from a multi-speaker mixture using cues that identify the target. Prior work has explored acoustic and textual cues, such as enrollment audio and different types of speaker descriptors, but typically relies on a single cue family. Yet no single cue is both broadly available and reliably discriminative. Moreover, most TSE systems adopt mask-prediction models, which are prone to audible artifacts and generalize poorly across acoustic conditions. To address these limitations, we propose PromptSep-Speaker, a generative TSE model based on latent diffusion that unifies four interchangeable cues: speaker temporal order, transcript, enrollment audio, and speaker caption. The generative formulation delivers higher separation quality and multi-modal prompting improves applicability across scenarios. To fully exploit each cue, we further introduce a frozen WavLM representation with an auxiliary alignment loss that supervises query-mixture cross-attention without extra labeling. Across four evaluation sets and a listening study, PromptSep-Speaker achieves state-of-the-art performance over TSE and universal source separation baselines, improving separation accuracy and quality across cue modalities.
Primary: Adobe Research
All Institutions: Adobe Research, Massachusetts Institute of Technology
PromptSep-Speaker introduces a generative latent-diffusion framework for target speaker extraction that unifies four interchangeable cues (speaker ID, transcript, enrollment, and caption) and employs a novel auxiliary alignment loss to supervise cross-attention between text cues and WavLM features, achieving state-of-the-art performance in both objective metrics and subjective listening tests across diverse acoustic conditions.
The paper proposes PromptSep-Speaker, a generative target speaker extraction (TSE) model based on latent diffusion. The core methodological contribution is the unification of four distinct cue modalities (speaker temporal order, transcript, enrollment audio, and speaker caption) within a single framework. A significant technical innovation is the "alignment module," which uses a frozen WavLM representation of the mixture to cross-attend to text-based cues (transcript/speaker ID). This is coupled with a novel auxiliary alignment loss that supervises the cross-attention maps to ensure the model correctly localizes the target speaker in the time domain without requiring extra labeling. The use of a DiT (Diffusion Transformer) backbone fine-tuned from a pre-trained source separation model is a strong architectural choice, leveraging prior knowledge of audio generation. The integration of semantic features via WavLM to guide attention is a clever way to bridge the gap between high-level linguistic cues and low-level acoustic separation.
The evaluation is rigorous and comprehensive. The authors test on four distinct datasets (CapSpeech-Mix, Libri2Mix, LibriTTS-R-Mix, and an internal set), covering both in-domain and out-of-domain scenarios. They compare against a diverse set of baselines, including mask-based TSE systems (DAE-TSE, SoloSpeech, WeSep) and generative universal source separators (PromptSep, FlowSep, SAM-Audio). The results show consistent state-of-the-art performance across SI-SDR, mel-L2, DNSMOS, and F1 decision error metrics. The inclusion of a subjective listening test with 380 subjects and 4,515 submissions adds significant weight to the objective metrics, confirming that the generative approach yields better perceptual quality and speaker accuracy than mask-based alternatives. The ablation studies effectively demonstrate the necessity of the auxiliary alignment loss and the specific direction of cross-attention.
The paper provides a good level of detail regarding the architecture (DiT, DAC-VAE, WavLM) and training hyperparameters (learning rate, sampler steps). However, the reliance on a "large internal collection of licensed data" for training limits full reproducibility by external groups. The specific composition of the 4.8M speech samples and the simulation pipeline details are high-level. While the demo page is provided, the absence of a public code repository or pre-trained weights for the specific PromptSep-Speaker model (distinct from the base models) is a minor drawback, though common for industry-affiliated papers at the arXiv stage.
The model is computationally expensive due to the diffusion process, which may limit real-time application compared to mask-based models. The evaluation is primarily on English speech; generalization to other languages is not tested. The "internal set" is not publicly available, making it impossible for third parties to verify those specific results. The paper does not extensively discuss the failure modes of the generative approach, such as hallucinations or artifacts in complex acoustic environments, beyond general statements about quality.
This work has significant implications for the field of speech processing by demonstrating that generative models can outperform traditional discriminative (mask-based) approaches in target speaker extraction. The multimodal prompting capability makes the system more versatile and user-friendly, as users can provide whichever cue is most available (e.g., a transcript from a meeting or a voice sample). This could lead to more robust meeting transcription systems, assistive listening devices, and content creation tools. The alignment loss technique could be adapted to other multimodal generation tasks where precise temporal grounding of text to audio is required. PromptSep-Speaker introduces a generative latent-diffusion framework for target speaker extraction that unifies four interchangeable cues (speaker ID, transcript, enrollment, and caption) and employs a novel auxiliary alignment loss to supervise cross-attention between text cues and WavLM features, achieving state-of-the-art performance in both objective metrics and subjective listening tests across diverse acoustic conditions.
Speech models are increasingly used as an interface to AI systems, from transcription tools to voice agents. Existing audio backdoors require an attacker to introduce triggers at inference time. We study a different threat model in which the attacker compromises the model during fine-tuning but has no access to the victim's inference-time inputs. We introduce the Speaker-Activated Backdoor (SAB) setting, in which a target speaker's natural voice serves as the trigger. We develop a two-phase fine-tuning method that first makes speaker identity more accessible and then binds the target speaker to an attacker-chosen payload. Compared with a one-phase baseline with the same training budget, the two-phase method substantially improves selectivity and clean performance, and is especially important for content-dependent attacks. SAB achieves 90.2% and 97.0% backdoor success on Whisper-medium and Parakeet, with false-positive rates of 1.7% and 0%, respectively using as little as two minutes of real audio from the target speaker. We further demonstrate content-dependent semantic rewriting and downstream tool-use attacks, where malicious instructions are inserted into 96.9% of target-speaker transcripts. When these transcripts are passed to frontier LLM agents, GPT-4o and Claude (Opus, Sonnet, and Haiku), the injected instruction drives the agent to carry out the harmful action in up to 87.5% of cases. The attack generalizes to held-out recordings, multiple target speakers, and common acoustic degradations. These results establish speaker identity as a practical backdoor trigger that requires no inference-time attacker access and is invisible at the transcript layer.
Primary: Imperial College London
All Institutions: Imperial College London
The paper introduces Speaker-Activated Backdoors (SAB), a novel attack where a target speaker's natural voice triggers malicious behavior in speech models without inference-time attacker access. By proposing a two-phase fine-tuning method that enhances speaker discriminability and binds identity to payloads, the authors demonstrate high-success, selective backdoors across multiple ASR and audio-LLM architectures, including downstream harmful actions in tool-using agents, establishing speaker identity as a critical security vector in voice-based AI systems.
The paper introduces a novel threat model for speech models where the trigger is the speaker's identity rather than an injected acoustic perturbation. The proposed two-phase fine-tuning method (SAB) is technically sound: Phase 1 enhances speaker discriminability in the encoder representations, and Phase 2 binds the target speaker to a malicious payload while preserving clean performance on non-target speakers. The ablation studies effectively demonstrate that Phase 1 is crucial for content-dependent attacks (semantic rewriting) and improves selectivity in fixed-output attacks. The use of voice cloning (F5-TTS) to augment limited real target audio is a practical and effective strategy.
The experimental evaluation is comprehensive, covering multiple architectures (Whisper, Parakeet, Qwen-Omni) and attack types (transcript suppression, semantic rewriting, instruction injection). The inclusion of downstream tool-use attacks with frontier LLMs (GPT-4o, Claude) significantly strengthens the impact by demonstrating real-world consequences. The evaluation on held-out real recordings ensures the results are not artifacts of synthetic data. Robustness to acoustic degradation and defense mechanisms (clean fine-tuning, voice anonymization) are also thoroughly analyzed.
The paper provides detailed hyperparameters, dataset splits, and training procedures. However, the specific implementation details for the auxiliary speaker head and the exact prompts used for the LLM agents are not fully specified in the main text, though the appendix provides some additional details. The use of standard datasets (LibriSpeech, VoxCeleb2) and open-source models aids reproducibility.
The attack relies on the attacker having access to the fine-tuning pipeline, which is a standard assumption but still a barrier. The backdoor is susceptible to clean fine-tuning, which limits its persistence in dynamic environments. The evaluation is limited to English speakers and specific public figures, which may not generalize to all demographic groups or private individuals.
This work highlights a significant security vulnerability in the growing ecosystem of voice-activated AI agents. It demonstrates that speaker identity can be exploited as a control channel, posing risks to privacy and safety in applications like voice assistants and automated banking. The findings will likely prompt the development of new defenses focused on speaker-identity invariance and supply-chain security for speech models. The paper introduces Speaker-Activated Backdoors (SAB), a novel attack where a target speaker's natural voice triggers malicious behavior in speech models without inference-time attacker access. By proposing a two-phase fine-tuning method that enhances speaker discriminability and binds identity to payloads, the authors demonstrate high-success, selective backdoors across multiple ASR and audio-LLM architectures, including downstream harmful actions in tool-using agents, establishing speaker identity as a critical security vector in voice-based AI systems.
We introduce MixAgent, an agentic pipeline for music post-production that follows the workflow of audio engineers to build interpretable and editable mixing plans guided by a reference song. A grouping agent first routes related input tracks, such as drum recordings, into subgroups. A planning agent then uses retrieved examples to choose and order audio effects, such as volume control and reverb, for tracks, subgroups, and the full mix. To refine the mix, a detector identifies what needs correction, and a critic proposes numerical parameter adjustments from tracks to the full mix. In our experiments, prompting recent multimodal models to listen and revise often worsens the mix, making reliable correction a central challenge. We therefore build MixCritic, a dataset of 45,657 repair critiques derived from parameters fitted to professional mixes, and train the critic with supervised fine-tuning followed by preference optimization on rendered audio outcomes. On MedleyDB, learned refinement reduces error by 19% relative to the initial mix, with preference optimization reducing it by 8.4% over supervised training alone. Retrieved examples further improve effects selection, and a human listening test shows that MixAgent gets higher quality ratings than mixes produced by Qwen3-Omni, Gemini 3.8 Flash and GPT-4o-audio.
Primary: Sony Group Corporation
All Institutions: Rice University, Sony Corporate Technology Center America, Inc., Sony Group Corporation
MixAgent introduces a novel agentic pipeline for music post-production that combines retrieval-augmented planning with learned numerical refinement, achieving state-of-the-art results in reference-guided mixing and providing a robust framework for interpretable audio editing.
The paper proposes MixAgent, a multi-agent framework for reference-guided music mixing. The methodology is structured into three distinct phases: grouping (routing tracks to subgroups), planning (selecting and ordering audio effects via retrieval-augmented generation), and refinement (detecting and correcting parameter errors). A key methodological contribution is the construction of the MixCritic dataset, which pairs corrupted mixing states with repair critiques derived from professional mixes. The authors employ a two-stage training process for the critic: Supervised Fine-Tuning (SFT) followed by Direct Preference Optimization (DPO) on rendered audio outcomes. The system explicitly models the hierarchical nature of mixing (tracks -> subgroups -> full mix) and uses a bottom-up refinement strategy, which is shown to be superior to top-down or flat approaches. The use of a frozen planner with retrieved examples to guide effect selection is a practical and effective design choice that mitigates the tendency of LLMs to select suboptimal or overly short effect chains.
The experiments are rigorous and comprehensive. The authors evaluate on MedleyDB and AlbumDB, using non-overlapping excerpts to prevent data leakage. They compare against strong baselines, including recent multimodal LLMs (Qwen3-Omni, Gemini 3.8 Flash, GPT-4o-audio) under various prompting strategies (Single-pass, Self-Refine, Self-Critic). The results demonstrate that MixAgent outperforms these baselines in reconstruction error (MSS) and predicted quality metrics. A particularly insightful analysis is the comparison of feedback forms (perceptual, tool-specific, numerical), showing that numerical corrections trained via DPO yield the best performance. The human listening test, while small (11 participants), provides qualitative validation that the system produces mixes rated higher than LLM baselines, though still below professional human mixes. The ablation studies on retrieval, refinement order, and detector selection are thorough and provide clear insights into the system's components.
The paper provides high reproducibility. The authors release the code and demo on GitHub. They detail the architecture of the detector and critic, the training procedures (SFT and DPO), and the construction of the MixCritic dataset. The prompts used for the agents are included in the appendix. The use of standard datasets (MedleyDB, AlbumDB) and public models (Qwen, Gemini, GPT) further enhances reproducibility. The specific hyperparameters and training settings are listed in tables, allowing for faithful replication.
The primary limitation is the small size of the evaluation sets (20 MedleyDB clips, 14 AlbumDB clips), which limits statistical power and generalizability. The listening test with only 11 participants is also a weakness. The system relies on GRAFX-fitted parameters for supervision, which may not perfectly capture the intent of professional engineers. The audio effects library is limited to a subset of professional tools, and the plan is fixed after the planning stage, meaning the critic cannot add or remove effects. The computational cost of multiple LLM calls for mixing is also noted as a practical limitation.
This work has significant implications for the field of AI-assisted creative tools. By providing an interpretable and editable mixing plan, it addresses a key barrier to adoption of AI in professional audio workflows. The findings on the effectiveness of numerical feedback and preference optimization for audio parameter adjustment are broadly applicable to other domains where LLMs control numerical tools. The construction of the MixCritic dataset is a valuable resource for the community. The paper also highlights the challenges of evaluating audio quality with current metrics, pointing to the need for better perceptual evaluation methods. MixAgent introduces a novel agentic pipeline for music post-production that combines retrieval-augmented planning with learned numerical refinement, achieving state-of-the-art results in reference-guided mixing and providing a robust framework for interpretable audio editing.
This paper presents XingChen-UniFE, a unified large audio language model (LALM) for general speech restoration (GSR) and target speaker extraction (TSE). The proposed model employs complementary semantic and acoustic representations through a dual-stream encoder, where a self-supervised learning model provides hierarchical semantic features and a learnable acoustic branch captures fine-grained structures from complex spectra. A bidirectional cross-attention module further facilitates information exchange between the two representations. The fused representation conditions the LALM for discrete speech token generation with task and speaker information. Experiments demonstrate strong performance on both GSR and TSE and robust generalization to unseen degradation conditions.
Primary: China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.
All Institutions: China Telecom Artificial Intelligence Technology (Beijing) Co., Ltd.
The paper presents a unified LALM-based framework for speech restoration and speaker extraction that effectively combines semantic and acoustic features via bidirectional cross-attention. The technical contribution is solid, with strong empirical results demonstrating improved performance over existing generative baselines and good generalization capabilities, though the lack of subjective evaluation and multilingual testing limits the comprehensiveness of the validation.
The paper proposes XingChen-UniFE, a unified Large Audio Language Model (LALM) for General Speech Restoration (GSR) and Target Speaker Extraction (TSE). The core methodological contribution is a dual-stream encoder that fuses semantic representations from a frozen self-supervised learning (SSL) model (WavLM) with acoustic representations from a learnable branch processing complex STFT spectrograms. A bidirectional cross-attention module is introduced to facilitate explicit information exchange between these two streams before conditioning the LALM backbone. The LALM generates interleaved acoustic and semantic tokens autoregressively, which are then decoded by a neural audio codec (H-Codec). The approach addresses the limitation of SSL features lacking fine-grained acoustic details (like phase) by explicitly modeling complex spectra, which is a reasonable and technically sound design choice for restoration tasks.
The experiments are conducted on the Libri2Mix dataset for TSE and various simulated degradation conditions for GSR. The model is compared against strong baselines like LLaSE-G1, UniSE, and UniTok. The proposed model achieves state-of-the-art or competitive results across PESQ, STOI, SpeechBERTScore (SBS), and WER metrics. Notably, the paper demonstrates robust generalization to unseen degradation types (codec distortion, wind noise) that were excluded from training, which is a significant strength. Ablation studies confirm the complementary roles of the semantic and acoustic branches and the importance of the bidirectional fusion mechanism.
The paper provides sufficient implementation details, including model architecture dimensions (28 layers, 1024 hidden dim), training hyperparameters (400k steps, batch size 64, AdamW), and data composition (approx. 10k hours target speech, 405k hours noise). However, the specific codebase and pre-trained checkpoints are not linked in the provided text, and the reliance on specific internal or less common datasets (e.g., Emilia, CN-Celeb) may pose minor barriers for full reproduction without access to those resources.
The model is computationally expensive due to the large LALM backbone (502M trainable parameters) and the dual-stream encoder, which may limit real-time deployment on edge devices. The evaluation is primarily limited to English speech (Libri2Mix, LibriSpeech); performance on multilingual or non-English speech is not assessed. Additionally, the paper lacks subjective listening tests (MOS), relying solely on objective metrics, which may not fully capture perceptual quality nuances.
This work contributes to the trend of unifying speech enhancement tasks under generative frameworks. By leveraging LALMs, it opens avenues for more complex speech processing tasks that require semantic understanding alongside acoustic restoration. The demonstrated generalization to unseen degradations suggests potential for robust deployment in diverse real-world environments. The paper presents a unified LALM-based framework for speech restoration and speaker extraction that effectively combines semantic and acoustic features via bidirectional cross-attention. The technical contribution is solid, with strong empirical results demonstrating improved performance over existing generative baselines and good generalization capabilities, though the lack of subjective evaluation and multilingual testing limits the comprehensiveness of the validation.
Omni LLMs can naturally locate key audio and visual evidence and integrate information across modalities for audiovisual reasoning. Existing audio-visual reasoning benchmarks that mainly rely on multiple-choice questions usually score models solely based on answer correctness, making it difficult to assess whether a model truly forms a reasoning chain from evidence to the correct answer. In addition, textual or unimodal shortcuts may allow models to answer correctly without analyzing the necessary audio-visual evidence, further weakening the reliability of evaluation benchmarks. To address these issues, we introduce EvaOmni, an evidence-grounded benchmark containing 392 videos and 1,027 purely human-annotated question-answer pairs, covering 15 capabilities and three levels: retrieval and localization, basic understanding, and complex reasoning. We design a series of rigorous annotation procedures in which experts annotate audio and visual evidence and construct video-based distractors; humans review whether questions meet strict inspection criteria, and finally models filter textual and unimodal shortcuts. Beyond high-quality question-answer pairs, we require models to simultaneously return the answer and supporting evidence intervals in order to evaluate the complete reasoning chain. We propose the Coverage Grounded Score (C.G.) to measure how much annotated evidence a model identifies, and the Bottleneck Grounded Score (B.G.) to penalize poor temporal localization in either modality; both are incorporated into answer correctness. Across 7 open-weight and 17 proprietary Omni LLMs, Gemini3.7-Flash achieved the highest overall accuracy (79.4%), C.G. (70.8%), and B.G. (35.1%). At Level 3, it reached 79.1% accuracy, but its B.G. was only 12.3%. These results show that even the most advanced models still have room for development in grounded audio-visual reasoning.
Primary: Shanghai Jiao Tong University
All Institutions: Shanghai Jiao Tong University, Shanghai Innovation Institute, Inner Mongolia University of Technology, Fudan University, Beijing Institute of Technology, Tongji University, The Hong Kong University of Science and Technology (Guangzhou), Tsinghua University, The University of Hong Kong, Tencent Youtu Lab
EvaOmni introduces an evidence-grounded benchmark for audio-visual reasoning that requires models to provide supporting evidence intervals, revealing significant gaps in temporal localization capabilities of current Omni LLMs. The paper makes a strong contribution by addressing the limitations of answer-only evaluation in multimodal benchmarks, proposing rigorous annotation procedures to mitigate shortcuts, and introducing novel metrics (C.G. and B.G.) that assess both answer correctness and evidence grounding. The extensive evaluation across 24 models demonstrates that while top models achieve high accuracy, their ability to precisely localize supporting evidence in both audio and visual modalities remains limited, providing valuable insights for future development of Omni LLMs.
The paper proposes EvaOmni, a benchmark that shifts the evaluation paradigm from answer-only correctness to evidence-grounded reasoning. The core methodological contribution is the requirement for models to output timestamped evidence intervals for both audio and visual modalities. The authors introduce two novel metrics: Coverage Grounded Score (C.G.), which measures the overlap of predicted evidence with ground truth, and Bottleneck Grounded Score (B.G.), which penalizes the weaker modality's localization accuracy. The annotation pipeline is rigorous, involving expert annotation, cross-validation, and model-assisted filtering to remove unimodal or textual shortcuts. This approach effectively addresses the "shortcut" problem in existing benchmarks where models can answer correctly without processing the full audio-visual context.
The evaluation is extensive, covering 24 Omni LLMs (7 open-weight, 17 proprietary) across 1,027 question-answer pairs. The results clearly demonstrate that while top models like Gemini3.7-Flash achieve high accuracy (79.4%), their grounding scores (B.G. 35.1%) are significantly lower, revealing a gap in precise temporal localization. The ablation studies on input modalities (Audio vs. ASR, Video vs. Audio+Video) provide valuable insights into how models utilize different cues. The comparison with human experts further highlights the difficulty of the task, as humans achieve much higher grounding scores. The cost-efficiency analysis adds practical value for model selection.
The paper provides detailed descriptions of the annotation process, metric definitions, and experimental setup. The authors state that data, annotations, and evaluation framework are fully released. However, specific URLs for the dataset or code are not explicitly provided in the text snippet, which slightly hinders immediate reproducibility verification, though the claim of release is strong. The prompt templates and preprocessing details are described in the appendix, aiding reproducibility.
The benchmark is limited to videos of 1-12 minutes, which may not capture reasoning over longer time horizons. The exclusion of hard-coded subtitles, while preventing OCR shortcuts, limits the diversity of real-world video content. The reliance on multiple-choice questions, even with grounding, may not fully capture open-ended reasoning capabilities. The human baseline allows replay, which is not possible for models, creating a slight asymmetry in comparison conditions.
This work has significant impact on the evaluation of multimodal LLMs. By introducing grounded metrics, it provides a more reliable way to assess whether models truly understand audio-visual content or are relying on shortcuts. This can guide future model development towards better temporal localization and cross-modal integration. The findings that even state-of-the-art models struggle with precise grounding highlight a key area for future research in Omni LLMs. EvaOmni introduces an evidence-grounded benchmark for audio-visual reasoning that requires models to provide supporting evidence intervals, revealing significant gaps in temporal localization capabilities of current Omni LLMs. The paper makes a strong contribution by addressing the limitations of answer-only evaluation in multimodal benchmarks, proposing rigorous annotation procedures to mitigate shortcuts, and introducing novel metrics (C.G. and B.G.) that assess both answer correctness and evidence grounding. The extensive evaluation across 24 models demonstrates that while top models achieve high accuracy, their ability to precisely localize supporting evidence in both audio and visual modalities remains limited, providing valuable insights for future development of Omni LLMs.
Voice style conversion (VSC) aims to transform a source utterance to match the timbre, accent, and emotion of a target voice while preserving linguistic content. StyleStream 1.0 introduced the first streamable zero-shot VSC system with state-of-the-art conversion quality, but remained limited in two ways: approximately 1s end-to-end latency on consumer hardware, and the lack of a text-based interface for designing or editing target voices. We present StyleStream 2.0, a fast and controllable streaming VSC framework that addresses these limitations. To reduce latency, we train a few-step pixel MeanFlow model and further fine-tune it with autoregressive feedback inspired by Self Forcing, improving robustness under streaming inference. This reduces end-to-end latency to 520~ms on a consumer GPU, a 2x speedup over StyleStream 1.0, while maintaining conversion quality. To enable controllable style manipulation, we remove the mel-spectrogram context and instead route target style information through a compact style embedding. In this embedding space, we train a unified text-conditioned flow matching model that supports both text-based voice design, which maps natural language prompts to voice styles, and instruction-based voice editing, which modifies a specific attribute of a source utterance while preserving the rest. Experiments show that StyleStream 2.0 achieves the strongest target style fidelity on VSC, voice design, and voice editing, while remaining competitive on intelligibility and non-edited attribute preservation.
Primary: University of California, Berkeley
All Institutions: University of California, Berkeley
StyleStream 2.0 advances the field of streaming voice style conversion by achieving a 2x latency reduction through few-step pixel MeanFlow and autoregressive fine-tuning, while introducing a unified text-conditioned framework for controllable voice design and editing. The paper demonstrates that holistic style conversion can be made both fast and controllable, bridging the gap between high-quality offline conversion and real-time interactive applications.
The paper proposes a two-stage framework for streaming Voice Style Conversion (VSC) that effectively decouples content from style. The core methodological contribution is the replacement of the multi-step diffusion stylizer with a few-step pixel MeanFlow (pMF) model, which is further fine-tuned using an autoregressive feedback mechanism inspired by Self Forcing. This addresses the critical train-test mismatch in streaming inference where the model must condition on its own imperfect past outputs. Additionally, the authors introduce a unified text-conditioned DiT "Voice Controller" that operates in a compact style-embedding space, enabling both zero-shot voice design from text prompts and attribute-specific editing (age, gender, accent, emotion). The architectural choice to remove mel-spectrogram context and rely solely on a perceiver-resampler encoded style embedding is a significant design decision that facilitates the text-based control interface.
The experimental evaluation is comprehensive, covering reference-conditioned VSC, text-based voice design, and voice editing. The authors compare against strong baselines including CosyVoice 2.0/3.0, Vevo 2.0, and their own StyleStream 1.0. Results demonstrate a 2x latency reduction (520ms vs ~1s) while maintaining or improving style fidelity metrics (S-SIM, A-SIM, E-SIM). The subjective evaluation (MOS) is well-conducted with 40 participants, showing competitive naturalness and improved style consistency. The ablation studies effectively isolate the contributions of the autoregressive fine-tuning and the new style encoder, confirming their necessity for streaming robustness and style accuracy.
The paper provides substantial detail on the training data construction (CapSpeech-derived corpus), the specific models used for evaluation (VoxProfile, Whisper), and the inference settings (NFE, CFG scale, chunk sizes). The inclusion of the demo URL and detailed appendices on data generation and evaluation protocols enhances reproducibility. However, the specific hyperparameters for the DiT and MeanFlow training are not fully detailed in the main text, though referenced in appendices.
The system is evaluated primarily on English speech, and the generalization to other languages is not discussed. The text-based control is limited to four specific attributes (age, gender, accent, emotion), which may not cover the full spectrum of vocal qualities users might wish to control (e.g., pitch, speaking rate, specific vocal textures). The reliance on a frozen HuBERT encoder for content extraction may limit the system's ability to handle highly noisy or non-speech audio inputs.
This work has significant implications for real-time interactive speech systems, such as live dubbing, accessibility tools, and voice assistants. The ability to control voice style via natural language prompts lowers the barrier to entry for non-expert users and enables more intuitive interaction with speech synthesis systems. The reduction in latency makes streaming VSC practical for consumer hardware, potentially enabling new applications in gaming, virtual reality, and remote communication. StyleStream 2.0 advances the field of streaming voice style conversion by achieving a 2x latency reduction through few-step pixel MeanFlow and autoregressive fine-tuning, while introducing a unified text-conditioned framework for controllable voice design and editing. The paper demonstrates that holistic style conversion can be made both fast and controllable, bridging the gap between high-quality offline conversion and real-time interactive applications.
This paper presents a causal, training-free linker that allows a frozen streaming diarizer to track more speakers than it has output slots. Streaming end-to-end diarizers such as Streaming Sortformer emit a fixed number N of concurrent speaker slots. Above N speakers, they keep detecting speech but reuse slots across people, because N bounds how many speakers the model can express at once, not how many it will meet. The linker uses only the host's slot posteriors and the audio. It keeps the frames where one slot is active, splits each slot where the speaker changes, embeds each piece with an off-the-shelf speaker encoder, and matches it against a bank of speakers seen so far. It only relabels, so missed speech and false alarm are identical to the host's and every gain is a reduction in speaker confusion. On 1418 sessions from eight corpora, it reduces the 4-slot host's DER above its cap from 36.8% to 22.6%, and barely moves it below the cap. The same constants transfer to the host's 8-slot successor and to two streaming EEND hosts from a different model family. On courtroom hearings with up to 23 speakers it reduces cpWER from 82.7% to 24.4%.
Primary: Adalat AI
All Institutions: Adalat AI
The paper introduces a training-free, causal linker that enables frozen streaming diarizers to track more speakers than their fixed output slots by using global speaker embeddings to relabel segments. This approach significantly reduces speaker confusion in high-speaker-count scenarios without altering the host model's voice activity detection, offering a practical and effective solution for production-grade streaming diarization systems.
The paper proposes a causal, training-free post-processing module ("linker") for frozen streaming diarizers. The core insight is that the fixed number of output slots in models like Streaming Sortformer represents a concurrency limit, not a lifetime speaker limit. The method identifies "pure" frames where only one slot is active, extracts speaker embeddings using an off-the-shelf encoder (TitaNet), and maintains a global speaker bank to relabel segments. By strictly relabeling without altering voice activity boundaries, the method guarantees that Missed Speech and False Alarm rates remain identical to the host model, isolating the improvement to Speaker Confusion reduction. The approach is elegant in its simplicity and its adherence to the "frozen host" constraint, making it highly practical for production systems where retraining the base model is infeasible.
The experimental setup is robust, utilizing 1,418 sessions from eight diverse corpora, including Indic languages, earnings calls, and courtroom hearings. The evaluation correctly stratifies results by speaker count relative to the model's capacity cap. The results show a significant reduction in DER (36.8% to 22.6%) for sessions exceeding the 4-slot cap, with negligible impact on below-cap sessions. The comparison against a widened 8-slot host demonstrates that the linker provides greater gains than simply increasing the model's output channels, a strong argument for the proposed method. The inclusion of downstream transcription metrics (cpWER) further validates the practical utility of the diarization improvements.
The paper provides high reproducibility. It specifies the exact host model version (Sortformer v2.1), the speaker encoder (TitaNet-large), and the specific hyperparameters (link threshold 0.60, purity score threshold 60%). The use of standard scoring tools (pyannote 4.1) and the detailed description of the bootstrap confidence intervals allow for rigorous verification of the claims. The code is not explicitly linked in the text provided, but the methodological details are sufficient for re-implementation.
The primary limitation is the reliance on an off-the-shelf speaker encoder, which may not be optimal for all acoustic conditions (e.g., heavy overlap or noisy environments). The method is also limited by the host model's ability to detect speech; if the host misses speech, the linker cannot recover it. Additionally, the paper notes that the linker can sometimes over-count speakers in below-cap scenarios, indicating a potential trade-off in identity stability. The latency of the embedding extraction and matching process is not deeply analyzed, though it is described as running on CPU.
This work has significant impact on the deployment of streaming diarization systems in real-world applications where speaker counts are unpredictable, such as large meetings, courtrooms, and call centers. By decoupling speaker identity tracking from the fixed-capacity diarization model, it offers a scalable solution that can be retrofitted to existing production systems without retraining. This approach could be extended to other fixed-capacity sequence labeling tasks in speech processing. The paper introduces a training-free, causal linker that enables frozen streaming diarizers to track more speakers than their fixed output slots by using global speaker embeddings to relabel segments. This approach significantly reduces speaker confusion in high-speaker-count scenarios without altering the host model's voice activity detection, offering a practical and effective solution for production-grade streaming diarization systems.
Speech separation (SS) has witnessed substantial progress in recent years with the advent of neural network-based methods, achieving notable improvements on signal-level evaluation metrics. Existing speech separation systems are mostly optimized with signal-level objectives, and often yield outputs that are acceptable in waveform metrics but suboptimal for intelligibility-sensitive downstream tasks. To address this gap, we propose SLM-SS, a token-based generative speech separation framework that models multi-talker mixtures with speech language models. We first discretize clean speech into multi-codebook token sequences with neural codecs and serialize multi-speaker targets via Serialized Output Training (SOT). An encoder-decoder autoregressive speech language model predicts the zero-order sequence to establish speaker ordering and coarse linguistic content, and a non-autoregressive module then refines higher-order residual codebooks for efficient reconstruction. We further introduce a per-talker non-autoregressive strategy that simplifies high-order modeling and improves the stability. On Libri2Mix, SLM-SS with per-talker NAR achieves $6.92\%$ WER, approaching codec-only reconstruction. The WavTokenizer variant eliminates NAR refinement and achieves the highest mean subjective quality score among evaluated separation systems. Its UTMOS and DNSMOS advantages over discriminative baselines persist on unseen noisy and Chinese mixtures without fine-tuning.
Primary: Shanghai Jiao Tong University
All Institutions: Shanghai Jiao Tong University
[One sentence main contribution]. [Comprehensive analysis of the technical contribution, methodology, and significance to the field]. The paper introduces SLM-SS, a token-based generative speech separation framework that utilizes speech language models to predict neural codec tokens, achieving competitive intelligibility and subjective quality scores while offering a flexible architecture for variable-talker scenarios, though it faces challenges in computational efficiency and signal fidelity due to codec discretization limits.
The paper proposes SLM-SS, a generative speech separation framework that leverages Speech Language Models (SLMs) to predict discrete neural codec tokens instead of continuous waveforms. The core method involves discretizing clean speech into multi-codebook sequences using Encodec or WavTokenizer, serializing multi-speaker targets via Serialized Output Training (SOT), and using an autoregressive (AR) encoder-decoder to predict the zero-order codebook. A non-autoregressive (NAR) module then refines higher-order residual codebooks. The introduction of a "per-talker NAR" strategy is a key methodological contribution, simplifying high-order modeling by decoupling speaker transitions from residual refinement. The approach effectively bridges the gap between signal-level separation and intelligibility-oriented downstream tasks by operating in the discrete token space.
Experiments are conducted on the LibriMix dataset (2-mix and 3-mix). The paper provides a comprehensive comparison against strong discriminative baselines (BSRNN, SepFormer, TF-LocoFormer, MossFormer2). Results show that SLM-SS achieves competitive WER and subjective quality scores, with the WavTokenizer variant achieving the highest mean subjective quality score. The inclusion of subjective listening tests with 30 volunteers and statistical significance testing (Wilcoxon signed-rank) adds rigor. Ablations on temperature, number of codebooks, and model parameters are provided. The evaluation of variable talker counts (up to 4) demonstrates the framework's flexibility.
The paper provides detailed architectural descriptions (WavLM encoder, Whisper-style decoder), training hyperparameters (learning rate, beam size, epochs), and dataset preprocessing steps (VAD segmentation). However, specific code availability is not explicitly confirmed in the text, and some implementation details regarding the exact initialization of the NAR module and the specific "staged training" strategy are described at a high level. The reliance on specific pre-trained models (WavLM, Whisper, Encodec, WavTokenizer) aids reproducibility.
The primary limitation is the computational cost of autoregressive decoding, which limits real-time applicability compared to discriminative methods. The method is constrained by the quality of the neural codec; truncating Encodec to 8 codebooks introduces irreversible distortion that caps the upper bound of signal fidelity (PESQ). The paper acknowledges that WavTokenizer, while better for subjective quality, has higher WER, indicating a trade-off between perceptual quality and linguistic accuracy. The experiments are limited to the LibriMix dataset, which consists of clean speech mixtures; performance on noisy or reverberant environments is not extensively tested in the main results (though mentioned as a future direction or minor test).
This work contributes to the trend of using discrete token representations for speech processing, aligning with the broader shift towards LLM-based audio models. It offers a viable alternative to traditional discriminative separation for applications where intelligibility is paramount, such as multi-talker ASR. The framework's ability to handle variable numbers of talkers without explicit speaker count estimation is a significant practical advantage. [One sentence main contribution]. [Comprehensive analysis of the technical contribution, methodology, and significance to the field]. The paper introduces SLM-SS, a token-based generative speech separation framework that utilizes speech language models to predict neural codec tokens, achieving competitive intelligibility and subjective quality scores while offering a flexible architecture for variable-talker scenarios, though it faces challenges in computational efficiency and signal fidelity due to codec discretization limits.
We present Qwen-Audio-3.1-ASR, a family of audio-language models that builds upon its predecesor Qwen-Audio-3.0-ASR with improved speech recognition and expanded capabilities in speaker-diarized transcription and general audio understanding. The family retains support for 30 languages and 16 Chinese dialects, together with instruction-controlled capabilities including historical-context conditioning, hierarchical hotword customization, and native transcription polishing. Speaker-diarized ASR produces structured outputs containing speaker labels, timestamps, and recognized text, maintaining consistent speaker labels across turns and the representation of overlapping speech. The general audio understanding capabilities cover audio-, speech-, and music-captioning, temporal grounding and sound event detection, as well as audio question answering and multiple-choice reasoning. These tasks broaden the application scope from recognizing spoken words to interpreting speakers, acoustic events, and their temporal relations. To accommodate different deployment requirements, the family includes variants for asynchronous file transcription, low-latency streaming, interactive voice input, and non-streaming recognition, with speaker-diarized transcription and general audio understanding available in certain variants. Evaluations span conventional ASR, multilingual and dialectal speech, speaker-diarized transcription, and general audio understanding, demonstrating the capabilities of the Qwen-Audio-3.1-ASR family across transcription and broader audio analysis tasks.
Primary: Alibaba Group
All Institutions: Alibaba Token Foundry, Alibaba Group
[One sentence main contribution]. The paper presents Qwen-Audio-3.1-ASR, a unified audio-language model family that integrates high-accuracy multilingual speech recognition, speaker-diarized transcription, and general audio understanding through a novel training pipeline featuring On-Policy Distillation and asynchronous reinforcement learning.
The paper proposes Qwen-Audio-3.1-ASR, a family of audio-language models that integrates speech recognition, speaker diarization, and general audio understanding into a unified framework. The methodology is robust, leveraging a Mixture-of-Experts (MoE) LLM backbone initialized from Qwen3, paired with a pretrained audio encoder (BEST-RQ + AED). A significant methodological contribution is the "On-Policy Distillation" (OPD) strategy, which uses student-generated responses to provide token-level feedback from specialized teachers, addressing the distribution mismatch in conventional distillation. Additionally, the introduction of "FunVerl," an asynchronous reinforcement learning framework, allows for complex reward modeling (combining WER, DER, and temporal consistency) without the computational bottleneck of synchronous rollouts. The serialization of speaker-diarized output as a single autoregressive sequence with specific loss reweighting for speaker IDs is a practical and effective design choice.
The evaluation is comprehensive, covering multilingual ASR (30 languages), Chinese dialects (16 varieties), speaker-diarized transcription, and general audio understanding (captioning, QA, temporal grounding). The paper reports improvements over the previous generation (Qwen-Audio-3.0-ASR) and compares against strong baselines like Whisper and FireRedAudio. The use of specific metrics like cpWER (concatenated minimum-permutation word error rate) and DER for diarization is appropriate. However, the paper is a technical report, so some experimental details (e.g., specific dataset splits for the "general audio understanding" benchmarks) are less transparent than in a peer-reviewed journal article. The results demonstrate strong performance, particularly in the integrated capabilities of the "Next" variant.
Reproducibility is moderate. While the architecture and training stages are described in detail, the specific hyperparameters for the MoE layers, the exact composition of the 20 million hours of pretraining data, and the implementation details of the "FunVerl" framework are not fully disclosed. The code and models are likely available via the provided URL, which aids in practical reproduction, but the scientific reproducibility of the training pipeline from scratch is limited by the proprietary nature of the data and infrastructure.
The paper does not extensively discuss the computational cost of the MoE architecture during inference, which is a critical factor for deployment. The "general audio understanding" capabilities are limited to the "Next" variant, creating a fragmentation in the model family. Additionally, the reliance on synthetic data for certain tasks (like domain-entity augmentation) may introduce biases that are not fully analyzed. The lack of ablation studies on the specific contribution of the On-Policy Distillation versus standard RL is a minor gap.
This work has significant impact on the field of speech technology by demonstrating that a single model can handle not just transcription, but also speaker identification and semantic understanding of audio. This unification simplifies system architectures for applications like meeting assistants, media analysis, and accessibility tools. The open-sourcing of the model family (implied by the GitHub URL) will likely accelerate research in multimodal audio understanding. [One sentence main contribution]. The paper presents Qwen-Audio-3.1-ASR, a unified audio-language model family that integrates high-accuracy multilingual speech recognition, speaker-diarized transcription, and general audio understanding through a novel training pipeline featuring On-Policy Distillation and asynchronous reinforcement learning.
Speech models are increasingly used as an interface to AI systems, from transcription tools to voice agents. Existing audio backdoors require an attacker to introduce triggers at inference time. We study a different threat model in which the attacker compromises the model during fine-tuning but has no access to the victim's inference-time inputs. We introduce the Speaker-Activated Backdoor (SAB) setting, in which a target speaker's natural voice serves as the trigger. We develop a two-phase fine-tuning method that first makes speaker identity more accessible and then binds the target speaker to an attacker-chosen payload. Compared with a one-phase baseline with the same training budget, the two-phase method substantially improves selectivity and clean performance, and is especially important for content-dependent attacks. SAB achieves 90.2% and 97.0% backdoor success on Whisper-medium and Parakeet, with false-positive rates of 1.7% and 0%, respectively using as little as two minutes of real audio from the target speaker. We further demonstrate content-dependent semantic rewriting and downstream tool-use attacks, where malicious instructions are inserted into 96.9% of target-speaker transcripts. When these transcripts are passed to frontier LLM agents, GPT-4o and Claude (Opus, Sonnet, and Haiku), the injected instruction drives the agent to carry out the harmful action in up to 87.5% of cases. The attack generalizes to held-out recordings, multiple target speakers, and common acoustic degradations. These results establish speaker identity as a practical backdoor trigger that requires no inference-time attacker access and is invisible at the transcript layer.
Primary: Imperial College London
All Institutions: Imperial College London
The paper introduces Speaker-Activated Backdoors (SAB), a novel attack where a target speaker's natural voice triggers malicious behavior in speech models without inference-time attacker access. By proposing a two-phase fine-tuning method that enhances speaker discriminability and binds identity to payloads, the authors demonstrate high-success, selective backdoors across multiple ASR and audio-LLM architectures, including downstream harmful actions in tool-using agents, establishing speaker identity as a critical security vector in voice-based AI systems.
The paper introduces a novel threat model for speech models where the trigger is the speaker's identity rather than an injected acoustic perturbation. The proposed two-phase fine-tuning method (SAB) is technically sound: Phase 1 enhances speaker discriminability in the encoder representations, and Phase 2 binds the target speaker to a malicious payload while preserving clean performance on non-target speakers. The ablation studies effectively demonstrate that Phase 1 is crucial for content-dependent attacks (semantic rewriting) and improves selectivity in fixed-output attacks. The use of voice cloning (F5-TTS) to augment limited real target audio is a practical and effective strategy.
The experimental evaluation is comprehensive, covering multiple architectures (Whisper, Parakeet, Qwen-Omni) and attack types (transcript suppression, semantic rewriting, instruction injection). The inclusion of downstream tool-use attacks with frontier LLMs (GPT-4o, Claude) significantly strengthens the impact by demonstrating real-world consequences. The evaluation on held-out real recordings ensures the results are not artifacts of synthetic data. Robustness to acoustic degradation and defense mechanisms (clean fine-tuning, voice anonymization) are also thoroughly analyzed.
The paper provides detailed hyperparameters, dataset splits, and training procedures. However, the specific implementation details for the auxiliary speaker head and the exact prompts used for the LLM agents are not fully specified in the main text, though the appendix provides some additional details. The use of standard datasets (LibriSpeech, VoxCeleb2) and open-source models aids reproducibility.
The attack relies on the attacker having access to the fine-tuning pipeline, which is a standard assumption but still a barrier. The backdoor is susceptible to clean fine-tuning, which limits its persistence in dynamic environments. The evaluation is limited to English speakers and specific public figures, which may not generalize to all demographic groups or private individuals.
This work highlights a significant security vulnerability in the growing ecosystem of voice-activated AI agents. It demonstrates that speaker identity can be exploited as a control channel, posing risks to privacy and safety in applications like voice assistants and automated banking. The findings will likely prompt the development of new defenses focused on speaker-identity invariance and supply-chain security for speech models. The paper introduces Speaker-Activated Backdoors (SAB), a novel attack where a target speaker's natural voice triggers malicious behavior in speech models without inference-time attacker access. By proposing a two-phase fine-tuning method that enhances speaker discriminability and binds identity to payloads, the authors demonstrate high-success, selective backdoors across multiple ASR and audio-LLM architectures, including downstream harmful actions in tool-using agents, establishing speaker identity as a critical security vector in voice-based AI systems.
A cover song re-renders an existing piece, most commonly by preserving its tonal content -- melody and chord progression -- while reshaping other musical attributes. This suggests a natural formulation for generative cover synthesis with pretrained music language models: explicitly control tonal content, while specifying lyrics and style through text. Existing approaches condition on frame-level features, tightly coupling generation to the reference's timing and structure, and limiting flexibility such as tempo variation, structural rearrangement, or partial conditioning. To this end, we present MuLaCover, a cover generation model that conditions a pretrained text-to-song foundation model on an editable symbolic lead sheet. Our design enables alignment-free control: generated outputs preserve the tonal signature of the source without frame-level alignment, allowing flexible timing, structure, and segment-level generation. We further introduce a training curriculum with partially mismatched audio-symbolic pairs, improving diversity and robustness. We evaluate MuLaCover with objective metrics, standard subjective ratings, and in-depth expert interviews that provide fine-grained diagnostic insights. Experiments show that MuLaCover achieves state-of-the-art performance among open-source systems and is competitive with leading commercial models such as Suno-v5.5. Audio samples and code are available at our project page: https://mulalabs.ai/mulacover/
Primary: MBZUAI
All Institutions: MBZUAI, Peking University, MuLa Labs, Vera Praxis Lab, The Chinese University of Hong Kong, Nanjing University of the Arts
MuLaCover introduces a flexible cover song generation framework by conditioning a text-to-song model on BPM-normalized symbolic lead sheets via gated cross-attention. The paper demonstrates that decoupling tonal content from timing through symbolic representation and cross-cover training enables superior controllability and diversity compared to frame-level conditioning, achieving performance competitive with leading commercial systems.
The paper proposes MuLaCover, a system for cover song generation that conditions a pretrained text-to-song foundation model (HeartMuLa) on a symbolic lead sheet (melody + chords) rather than frame-level audio features. The core methodological contribution is the use of BPM-normalized symbolic representations to decouple tonal content from timing, enabling flexible tempo and structural changes. The model injects this condition via a Gated Adaptive Cross-Attention (GACA) module into the last layers of the backbone. A key innovation is the three-stage training curriculum, specifically the "cross-cover" finetuning stage where tonal content from one rendition is paired with lyrics/style from another, effectively teaching the model to disentangle content from style. The ablation studies rigorously validate the necessity of BPM normalization, the gate mechanism, and the cross-cover training strategy.
The evaluation is comprehensive, utilizing objective metrics (AudioBox-Aesthetics, SongEval, PER, MuQ similarity, CSI accuracy) and subjective evaluations (listening study with 22 participants, expert interviews). The paper compares against strong baselines including open-source systems (SongEcho, ACE-Step-1.5) and a leading commercial model (Suno-v5.5). Results show MuLaCover achieves state-of-the-art performance among open-source systems and is competitive with Suno-v5.5. The partial-reference cover task (inpainting/outpainting) is also evaluated, demonstrating the flexibility of the alignment-free approach. The expert interview protocol provides qualitative insights into why certain systems perform better on specific musical attributes.
The paper provides detailed descriptions of the architecture, training stages, and data pipeline. It mentions the use of specific models for transcription (YourMT3, etc.) and data synthesis. While code is stated to be available at the project page, the reliance on synthetic data from commercial providers and specific internal pipelines (like the cross-cover pairing logic) might limit full external reproduction without access to those specific data assets. However, the methodological details are sufficient for implementation.
The system relies on the quality of the symbolic transcription models for melody and chords; errors in transcription will propagate to the generation. The training data is synthetic, which may limit the model's ability to handle the nuances of human-recorded covers with complex timbral variations. The "cover" definition is fixed at the lead-sheet level, which may not capture all artistic interpretations of a cover (e.g., jazz re-harmonization).
This work advances controllable music generation by providing a flexible, alignment-free method for conditioning on musical structure. It bridges the gap between symbolic music representation and audio generation, offering a practical tool for music production and co-creation. The approach could be extended to other domains where structural control is needed without strict temporal alignment. MuLaCover introduces a flexible cover song generation framework by conditioning a text-to-song model on BPM-normalized symbolic lead sheets via gated cross-attention. The paper demonstrates that decoupling tonal content from timing through symbolic representation and cross-cover training enables superior controllability and diversity compared to frame-level conditioning, achieving performance competitive with leading commercial systems.
We introduce Long Listening Thoughts (LLT), a framework for constructing post-training datasets that improves audio reasoning by pairing deliberate perceptual grounding with explicit cognitive behaviors. On audio understanding and reasoning benchmarks, large audio-language models (LALMs) are limited primarily by perceptual errors, yet existing post-training datasets rely on audio descriptions of limited diversity. Through a four-stage pipeline, LLT (1) builds multi-aspect dense audio descriptions spanning semantic, acoustic, and temporal information, (2) generates more challenging questions, (3) synthesizes incorrect reasoning traces, and (4) extends short reasoning traces into long ones via a thought-continuation strategy that elicits cognitive behaviors such as verification and backtracking. Using LLT, we synthesize 1.6M reasoning traces and train Qwen2.5-Omni-7B with a two-stage pipeline of supervised fine-tuning (SFT) followed by direct preference optimization (DPO). This improves accuracy on MMAU Test-mini (71.5 to 78.2), MMAR (56.7 to 63.7), and MMSU (60.6 to 66.0).
Primary: Carnegie Mellon University
All Institutions: Carnegie Mellon University, NVIDIA
The paper introduces a scalable four-stage pipeline for constructing post-training datasets that enhance audio reasoning by combining dense multi-aspect perceptual grounding with explicit cognitive behaviors like verification and backtracking, demonstrating significant improvements in LALMs through a two-stage SFT+DPO training strategy.
The paper proposes a four-stage data construction pipeline (LLT) to improve audio reasoning in Large Audio-Language Models (LALMs). The core innovation lies in decoupling perceptual grounding from cognitive refinement. Stage 1 creates dense, multi-aspect audio descriptions (semantic, acoustic via spectrogram-VLM, temporal via specialized detectors) to overcome the limited perceptual diversity of existing datasets. Stage 2 augments question difficulty. Stage 3 synthesizes plausible incorrect reasoning traces using an LLM, rather than relying on model-sampled errors which lack diversity. Stage 4 employs a "thought-continuation" strategy to extend short traces into long ones with explicit cognitive behaviors (verification, backtracking). The training strategy is a two-stage process: SFT on correct and refined traces, followed by DPO using preference pairs that contrast incorrect-to-correct traces. This separation is methodologically sound, as the authors demonstrate that including incorrect traces directly in SFT degrades performance due to the model's inability to distinguish intermediate errors from correct reasoning under standard supervision.
The experiments are extensive and well-controlled. The authors validate the necessity of multi-aspect descriptions by showing gains when feeding them to a text-only LLM. They ablate the source of incorrect traces (synthetic vs. model-sampled), showing synthetic traces provide better coverage and distributional separation for DPO. The two-stage training is rigorously compared against single-stage SFT variants. Results show consistent improvements over the Qwen2.5-Omni-7B baseline on MMAU, MMAR, and MMSU. Generalization is tested on different source datasets (AudioMCQ) and different backbones (Qwen2-Audio, Qwen3-Omni), showing the pipeline's robustness. The analysis of cognitive triggers confirms that the model actually engages in error correction (8.7% true corrections) rather than just generating longer text.
High. The authors release the dataset (1.6M traces), trained models, and the full data construction pipeline. All components (Qwen3-Omni-Captioner, Whisper, pyannote, BEATs, etc.) are open-source. Prompts and implementation details are provided in the appendix. The use of standard benchmarks and clear ablation protocols enhances reproducibility.
The pipeline relies heavily on the quality of the initial audio CoT dataset (AF-Think) and the capabilities of the teacher LLMs (Qwen3-Next-80B). The 30-second audio limit restricts applicability to long-form audio, though some transfer is observed. The automated verification of traces may miss subtle logical errors, although human inspection mitigates this. The computational cost of generating 1.6M traces with multiple specialized models is significant.
This work provides a scalable framework for enhancing audio reasoning in LALMs, which is crucial for applications in assistive technology, audio surveillance, and multimodal agents. By explicitly teaching cognitive behaviors like verification and backtracking, it moves beyond simple pattern matching to more robust reasoning. The release of the dataset and pipeline will likely accelerate research in audio post-training. The paper introduces a scalable four-stage pipeline for constructing post-training datasets that enhance audio reasoning by combining dense multi-aspect perceptual grounding with explicit cognitive behaviors like verification and backtracking, demonstrating significant improvements in LALMs through a two-stage SFT+DPO training strategy.
Target speaker extraction (TSE) isolates a desired speaker from a multi-speaker mixture using cues that identify the target. Prior work has explored acoustic and textual cues, such as enrollment audio and different types of speaker descriptors, but typically relies on a single cue family. Yet no single cue is both broadly available and reliably discriminative. Moreover, most TSE systems adopt mask-prediction models, which are prone to audible artifacts and generalize poorly across acoustic conditions. To address these limitations, we propose PromptSep-Speaker, a generative TSE model based on latent diffusion that unifies four interchangeable cues: speaker temporal order, transcript, enrollment audio, and speaker caption. The generative formulation delivers higher separation quality and multi-modal prompting improves applicability across scenarios. To fully exploit each cue, we further introduce a frozen WavLM representation with an auxiliary alignment loss that supervises query-mixture cross-attention without extra labeling. Across four evaluation sets and a listening study, PromptSep-Speaker achieves state-of-the-art performance over TSE and universal source separation baselines, improving separation accuracy and quality across cue modalities.
Primary: Adobe Research
All Institutions: Adobe Research, Massachusetts Institute of Technology
PromptSep-Speaker introduces a generative latent-diffusion framework for target speaker extraction that unifies four interchangeable cues (speaker ID, transcript, enrollment, and caption) and employs a novel auxiliary alignment loss to supervise cross-attention between text cues and WavLM features, achieving state-of-the-art performance in both objective metrics and subjective listening tests across diverse acoustic conditions.
The paper proposes PromptSep-Speaker, a generative target speaker extraction (TSE) model based on latent diffusion. The core methodological contribution is the unification of four distinct cue modalities (speaker temporal order, transcript, enrollment audio, and speaker caption) within a single framework. A significant technical innovation is the "alignment module," which uses a frozen WavLM representation of the mixture to cross-attend to text-based cues (transcript/speaker ID). This is coupled with a novel auxiliary alignment loss that supervises the cross-attention maps to ensure the model correctly localizes the target speaker in the time domain without requiring extra labeling. The use of a DiT (Diffusion Transformer) backbone fine-tuned from a pre-trained source separation model is a strong architectural choice, leveraging prior knowledge of audio generation. The integration of semantic features via WavLM to guide attention is a clever way to bridge the gap between high-level linguistic cues and low-level acoustic separation.
The evaluation is rigorous and comprehensive. The authors test on four distinct datasets (CapSpeech-Mix, Libri2Mix, LibriTTS-R-Mix, and an internal set), covering both in-domain and out-of-domain scenarios. They compare against a diverse set of baselines, including mask-based TSE systems (DAE-TSE, SoloSpeech, WeSep) and generative universal source separators (PromptSep, FlowSep, SAM-Audio). The results show consistent state-of-the-art performance across SI-SDR, mel-L2, DNSMOS, and F1 decision error metrics. The inclusion of a subjective listening test with 380 subjects and 4,515 submissions adds significant weight to the objective metrics, confirming that the generative approach yields better perceptual quality and speaker accuracy than mask-based alternatives. The ablation studies effectively demonstrate the necessity of the auxiliary alignment loss and the specific direction of cross-attention.
The paper provides a good level of detail regarding the architecture (DiT, DAC-VAE, WavLM) and training hyperparameters (learning rate, sampler steps). However, the reliance on a "large internal collection of licensed data" for training limits full reproducibility by external groups. The specific composition of the 4.8M speech samples and the simulation pipeline details are high-level. While the demo page is provided, the absence of a public code repository or pre-trained weights for the specific PromptSep-Speaker model (distinct from the base models) is a minor drawback, though common for industry-affiliated papers at the arXiv stage.
The model is computationally expensive due to the diffusion process, which may limit real-time application compared to mask-based models. The evaluation is primarily on English speech; generalization to other languages is not tested. The "internal set" is not publicly available, making it impossible for third parties to verify those specific results. The paper does not extensively discuss the failure modes of the generative approach, such as hallucinations or artifacts in complex acoustic environments, beyond general statements about quality.
This work has significant implications for the field of speech processing by demonstrating that generative models can outperform traditional discriminative (mask-based) approaches in target speaker extraction. The multimodal prompting capability makes the system more versatile and user-friendly, as users can provide whichever cue is most available (e.g., a transcript from a meeting or a voice sample). This could lead to more robust meeting transcription systems, assistive listening devices, and content creation tools. The alignment loss technique could be adapted to other multimodal generation tasks where precise temporal grounding of text to audio is required. PromptSep-Speaker introduces a generative latent-diffusion framework for target speaker extraction that unifies four interchangeable cues (speaker ID, transcript, enrollment, and caption) and employs a novel auxiliary alignment loss to supervise cross-attention between text cues and WavLM features, achieving state-of-the-art performance in both objective metrics and subjective listening tests across diverse acoustic conditions.
The text contains content addressed to automated reviewers. Its score has been penalised, and the ranking below should not be trusted.
[system_override] "at once. Tool results are treated as evidence and never as new instructions, a repeated identical call returns the earlier receipt inst"
Text addressed to an automated reviewer was found in this paper. The impact score shown has already been penalised for this, but treat the ranking with suspicion.
[system_override] "at once. Tool results are treated as evidence and never as new instructions, a repeated identical call returns the earlier receipt inst"
Interactive models increasingly hand complex reasoning and tool use to a slow brain in the background, leaving a realtime front model free to sustain full-duplex dialogue. Yet the user may refine, cancel or replace a request while its task runs, so a plan or result that made sense a moment ago goes stale. Each new utterance must therefore be read against the work in flight, but the front model follows only the conversation and the back end sees only its own task. Coordination deserves a brain of its own, one that sees both. We introduce Jarvis, a coordination harness where models judge and code enforces. Its typed decision model, JEV, reads each turn against the live tasks and picks among operations built from a versioned task record. Code checks every choice against the current state before applying it, so no decision can act on a task version the user has replaced. Jarvis also rechecks plans before they execute and results before they are spoken, and counts an answer as delivered only once the user has heard it in full. We build two instances of this design. Jarvis-Audio runs unchanged on three full-duplex speech models. Jarvis-Omni adds vision and rechecks the evidence behind an answer before and during playback, since a scene can change while nobody speaks. With the front model and back end held fixed, replacing Qwen Audio Agent with Jarvis raises tau-Voice task success more than sixfold, from 29 to 192 of 278 tasks, and the gains hold on every benchmark and front model. Against published systems, Jarvis-Audio reaches 69% Pass1 on Full-Duplex-Bench v3 with an open 9B front model, versus 60% for GPT-Realtime. These results suggest that how the two brains are coordinated matters as much as how strong they are.
Primary: Tencent Hunyuan
All Institutions: The Chinese University of Hong Kong, Shenzhen, Tencent Hunyuan, Amphion Technology Co., Ltd.
The paper introduces a robust coordination harness for full-duplex interactive models that resolves race conditions and stale-state errors by employing a dedicated typed decision model (JEV) to select from runtime-generated, versioned operations, thereby ensuring that background tasks remain synchronized with live user dialogue. This architectural innovation, validated by significant performance gains over existing agents like Qwen Audio Agent across multiple benchmarks, establishes a new standard for reliability in agentic speech systems.
The paper proposes "Jarvis," a coordination harness for full-duplex interactive models that separates the "fast brain" (real-time speech model) from the "slow brain" (background agent) and introduces a dedicated "coordination brain" (JEV). The core methodological contribution is the "models judge, code enforces" paradigm. Instead of allowing the LLM to generate free-form actions or task updates, the runtime constructs a set of typed, versioned operations (e.g., REFINE:task_id:version) based on the current state of a "Task Ledger." The JEV model (a typed judgment model) selects from these pre-constructed options. This ensures that no action can reference a non-existent task or a superseded version, effectively eliminating race conditions and stale-state errors. The system also implements a "Native Gate" trained via GRPO on the front model's control tokens to decide when to delegate, and a receipt-based delivery mechanism that distinguishes between a task being "done" and "heard" by the user. The extension to vision (Jarvis-Omni) treats visual evidence as time-bounded state, rechecking validity before and during playback.
The experimental evaluation is rigorous and comprehensive. The authors isolate the impact of the harness by holding the front model (Venus-Audio, MiniCPM-o 4.5, VoiceChat-11B) and back-end fixed, showing that replacing the baseline Qwen Audio Agent (QAA) with Jarvis yields massive improvements (e.g., 29 to 192 tasks solved on tau-Voice). They provide detailed ablations on the "race suite" (80 scripts x 3 runs) demonstrating that each invariant (version checking, plan freshness, playback receipts) prevents specific failure modes. The comparison against published systems (GPT-Realtime, etc.) on Full-Duplex-Bench v3 shows competitive or superior performance. The vision experiments confirm that removing visual evidence significantly degrades performance, validating the need for the evidence timeline.
The paper provides high reproducibility standards. It explicitly states that code is available at the provided GitHub URL. It details the training data for the Native Gate (264 authored scenarios, GRPO training details, LoRA rank 8). It specifies the models used (DeepSeek-V4-Flash for workers, Qwen3-ASR, etc.) and the benchmark protocols. The "race suite" for testing concurrency is a valuable contribution to reproducibility in this domain.
The system relies on a specific "front model" architecture (Venus-Audio/MiniCPM-o style) with native control tokens for the gate; adapting to models without such tokens might require different gating strategies. The "JEV" model is a separate component, adding latency (though reported as <1s median). The vision component is tested primarily on game scenarios (CS2, LoL, PvZ), so generalization to other visual domains (e.g., robotics, AR) is not fully explored. The reliance on a "slow brain" (DeepSeek-V4-Flash) means the system inherits the cost and latency of large LLM inference for complex tasks.
This work addresses a critical bottleneck in the deployment of real-time AI assistants: the coordination between real-time interaction and background computation. By formalizing the "coordination brain" and enforcing state consistency via code, it provides a robust template for building reliable agentic systems. The approach has implications for any system where real-time dialogue intersects with asynchronous tool use or long-running tasks, including customer service bots, smart home assistants, and embodied AI. The paper introduces a robust coordination harness for full-duplex interactive models that resolves race conditions and stale-state errors by employing a dedicated typed decision model (JEV) to select from runtime-generated, versioned operations, thereby ensuring that background tasks remain synchronized with live user dialogue. This architectural innovation, validated by significant performance gains over existing agents like Qwen Audio Agent across multiple benchmarks, establishes a new standard for reliability in agentic speech systems.
Human face-to-face conversation requires coordinating speech, gestures and expressions across both speaking and listening. We present ELLSA 2, an end-to-end full-duplex dialogue model that jointly optimizes speech interaction and streaming gesture-expression generation. An autoregressive, chunk-wise synthesizer combines pretrained acoustic features with distinct listening and speaking hidden states from the dialogue backbone and speech synthesizer, enabling continuous non-verbal generation even when the model is verbally silent. To translate dialogue understanding into vivid semantic gestures, we introduce contextual semantic calls and template guidance, separating whether to invoke an available semantic gesture from how to realize it. We also establish a curated data and evaluation pipeline for dialog-aware co-speech gesture generation and investigate reference-based multimodal large language models (MLLMs) as automatic judges. Experiments show that ELLSA 2 remains highly competitive with open-source full-duplex speech interaction models and outperforms cascaded baselines across dialog-aware gesture and expression metrics. Semantic template guidance further significantly improves semantic gesture quality. Strong system-level correlations with human judgments support these reference-based scores as useful preliminary metrics for full-duplex and semantic gesture assessment. Together, these results advance dialogue modeling from spoken responses toward coordinated multimodal interaction.
Primary: Tsinghua University
All Institutions: Tsinghua University, ByteDance
ELLSA 2 introduces a unified end-to-end framework for full-duplex speech interaction with coordinated gesture and expression generation, featuring a novel semantic template guidance mechanism that significantly improves the quality and appropriateness of non-verbal behaviors in dialogue.
The paper proposes ELLSA 2, a unified end-to-end framework that integrates full-duplex speech interaction with streaming gesture and facial expression generation. The core methodological contribution is the "contextual semantic call" mechanism, which decouples the decision of *whether* to perform a semantic gesture from *how* to realize it. This is achieved via a binary classification head on the LLM backbone's hidden states, which triggers the retrieval of a curated exemplar template to guide the autoregressive gesture generator. The architecture uses a chunk-wise autoregressive synthesizer that interleaves speech, listening, and motion tokens, allowing non-verbal behaviors to persist during verbal silence (listening states). The use of RVQ-VAE for motion tokenization and a curriculum training strategy to integrate modalities is technically sound and well-motivated.
The experimental setup is robust, introducing a new human-verified benchmark (138 clips) specifically for dialog-aware gesture generation, covering full-duplex behaviors and semantic gestures. The evaluation employs both objective metrics (prosodic alignment, activity ratio) and subjective human evaluation, along with MLLM-as-a-judge metrics. The results demonstrate that ELLSA 2 outperforms cascaded baselines and remains competitive with state-of-the-art full-duplex speech models. The semantic template guidance shows a significant improvement in semantic gesture quality (human score 2.014 to 2.711). The correlation analysis between MLLM judges and human annotators (0.853/0.919) provides strong evidence for the utility of the proposed automatic metrics.
The paper provides extensive implementation details, including tokenizer specifications, training hyperparameters, and data curation pipelines. The authors commit to releasing code, data, and model checkpoints upon acceptance. The detailed description of the semantic exemplar library construction and the evaluation protocols enhances reproducibility.
The system relies on a curated library of 3,227 exemplars for semantic gestures, which may limit the diversity of semantic actions compared to fully generative approaches. The evaluation, while improved, is still based on a relatively small test set (138 clips). The model's performance on long-horizon coherence of gestures over extended conversations is not deeply analyzed.
This work advances the field of human-computer interaction by moving beyond text or speech-only interfaces to fully embodied, multimodal agents. The proposed evaluation metrics and benchmark will be valuable for the community working on co-speech gesture generation and full-duplex dialogue systems. The integration of semantic understanding with physical gesture generation has implications for robotics, virtual avatars, and immersive communication. ELLSA 2 introduces a unified end-to-end framework for full-duplex speech interaction with coordinated gesture and expression generation, featuring a novel semantic template guidance mechanism that significantly improves the quality and appropriateness of non-verbal behaviors in dialogue.
As joint audio-video generation models become more capable, they can generate audio-visual scenes exhibiting three forms of complexity: multi-event composition, multi-layer audio mixture, and asymmetric cross-modal relations. However, current evaluations largely reuse metrics from earlier simple and unidirectional tasks. They typically evaluate an audio-video clip as a whole and rely on simple predefined relations, making them unsuitable for these increasingly complex scenes. We propose an event-centric evaluation framework that disentangles audio and visual events at the event level, analyzes each event independently, cross-validates relations across modalities, and aggregates the judgments into a coarse-to-fine, three-level consistency metric. The metric assesses global narrative harmony, event-level temporal synchronization, and fine-grained attribute alignment, providing a general formulation for audio-video consistency evaluation across diverse generation settings and cross-modal relations. We instantiate the metric as AVC-Reward, a trainable audio-video reward model. To evaluate the agreement between automatic judgments and human perception across generation paradigms and relation complexities, we construct AVC-Benchmark, which is primarily built from diverse positive and negative samples generated by various types of audio-video generation models. AVC-Reward achieves state-of-the-art alignment with human judgment among automatic methods, provides interpretable rationales that localize failures, and improves the audio-video consistency of a joint generation model when used as an online reinforcement-learning signal. Project page: https://tsuki0512.github.io/AVC-Reward.github.io/
Primary: Zhejiang University
All Institutions: Zhejiang University, HiThink Research, Fudan University, Lanzhou University
The paper introduces a novel event-centric, three-level consistency metric and a corresponding trainable reward model (AVC-Reward) that effectively evaluates and improves the audio-video consistency of joint generation models. By disentangling events and using a structured reasoning chain, it overcomes the limitations of traditional clip-level metrics, achieving state-of-the-art human alignment and demonstrating significant utility as a reinforcement learning signal for generative model optimization.
The paper proposes an event-centric evaluation framework for audio-video consistency, addressing the limitations of clip-level metrics in handling multi-event compositions, multi-layer audio mixtures, and asymmetric cross-modal relations. The core contribution is a three-level coarse-to-fine metric (Global Narrative Harmony, Event-level Temporal Synchronization, Fine-grained Attribute Alignment) implemented via a trainable reward model, AVC-Reward. The model is initialized from Qwen3-Omni-Thinking and trained using a two-stage process: Supervised Fine-Tuning (SFT) on structured reasoning data generated by a multi-LLM annotation pipeline, followed by Group Relative Policy Optimization (GRPO) to align scores with human preferences. The methodology is rigorous, introducing a 39-label taxonomy for human annotation to reduce subjectivity and a label-to-score calibration mechanism using LLMs to generate continuous targets for RL training. The five-stage reasoning protocol (blind audio analysis, blind video analysis, bidirectional verification, and final judgment) ensures that the model learns to disentangle events before scoring, which is a significant methodological improvement over black-box similarity metrics.
The authors construct AVC-Benchmark, the first benchmark primarily built from model-generated samples (A2V, V2A, JAVG) to evaluate human-alignment. AVC-Reward achieves state-of-the-art performance in pairwise accuracy, SRCC, and PLCC compared to baselines like SyncNet, ImageBind, and other learned evaluators. Crucially, the paper validates the utility of the reward model by using it in an online reinforcement learning setup (OmniNFT) for the LTX-2 generation model. This results in measurable improvements in external consistency metrics and human preference scores, demonstrating that the evaluator can effectively serve as a training signal to close the "generate-evaluate-optimize" loop. The inclusion of human evaluation with blinded randomized judgments further strengthens the experimental claims.
The paper provides detailed descriptions of the data construction pipeline, annotation protocol, and training hyperparameters. It references a project page and GitHub repository for prompts and code. The use of specific open-source models (Qwen3-Omni, LTX-2, OmniNFT) and public benchmarks (JavisBench, AV-Phys-Bench) enhances reproducibility. However, the reliance on specific internal models for some data generation and the complexity of the multi-LLM annotation harness may pose challenges for exact replication without access to the specific model versions and internal datasets.
The evaluation is heavily dependent on the quality of the initial LLM-based annotations for SFT, which may propagate biases from the annotator models (Gemini 2.5 Pro, Qwen3-VL-Max). The GRPO training relies on a relatively small set of human-calibrated samples (10k), which might limit the generalization of the reward model to unseen failure modes. Additionally, the computational cost of the five-stage reasoning process during inference could be high, potentially limiting its use in real-time applications or large-scale screening without distillation.
This work significantly advances the field of multimodal generation by providing a robust, interpretable, and trainable metric for audio-video consistency. It addresses a critical gap in evaluating modern joint generation models that produce complex scenes. The ability to use this metric as a reinforcement learning signal offers a practical pathway for improving generative models, making it highly impactful for both research and industrial applications in video and audio synthesis. The paper introduces a novel event-centric, three-level consistency metric and a corresponding trainable reward model (AVC-Reward) that effectively evaluates and improves the audio-video consistency of joint generation models. By disentangling events and using a structured reasoning chain, it overcomes the limitations of traditional clip-level metrics, achieving state-of-the-art human alignment and demonstrating significant utility as a reinforcement learning signal for generative model optimization.
A cascaded voice agent must decide, throughout each silence of its user, whether the user has finished their turn: answering too early cuts the user off, answering too late makes the agent slow. It often needs related decisions too: will the user soon go on, is speech during its own turn an interruption or a backchannel, is a sound speech at all. With open detectors, each extra decision costs another model or another call, and on the 14-language eot-bench they cut users off in more than a quarter of mid-turn pauses. Because they are trained on segmented utterances, this mismatch is the largest single source of error. We present VoxTurn, a 0.8B speech LLM that casts these decisions as multiple-choice questions, scores the hidden state of each option, and answers any set of them in one forward pass with a block-isolating attention mask, exactly as if asked one at a time. Trained on real conversational pauses with labels derived from their timing, and re-queried while a silence lasts, VoxTurn reaches 11.5% false cut-offs in English at a 300 ms latency budget, comparable to two commercial detectors, and a 14-language mean of 19.9%, the lowest among open models, at 8.3 ms per decision. In Pipecat and LiveKit it cuts off 17% and 9% of turns, and it ranks fifth of 24 in TurnBench end-of-turn detection without any TurnBench training data. The model, code and framework plugins are released.
Primary: University of Melbourne
All Institutions: University of Melbourne, National Computational Infrastructure (NCI Australia)
VoxTurn introduces a 0.8B speech LLM that answers multiple turn-taking questions in a single forward pass using a block-isolating attention mask, achieving commercial-competitive performance on end-of-turn detection with significantly improved efficiency and multilingual robustness.
The paper proposes VoxTurn, a 0.8B speech LLM based on Qwen3-ASR-0.6B, adapted for turn-taking decisions. The core methodological innovation is the "one-pass multi-question" inference mechanism using a block-isolating attention mask. This allows the model to answer multiple related questions (e.g., end-of-turn, interruption, backchannel, invalid input) in a single forward pass, mathematically equivalent to asking them sequentially but significantly more efficient. The training strategy is also notable: instead of using segmented utterances with static labels, the authors train on real conversational pauses, sampling decision times uniformly from within the pause and deriving labels from the timing of subsequent speech. This addresses the distribution mismatch between training (segmented) and deployment (streaming/real-time) scenarios. The use of a linear head on hidden states for multiple-choice scoring rather than text generation is a standard but effective adaptation for classification tasks in LLMs.
The experimental evaluation is comprehensive and rigorous. The authors evaluate on eot-bench (14 languages) and TurnBench (two-party conversations). Key results include 11.5% false cut-offs in English at 300ms latency, comparable to commercial detectors, and the best open-model performance on the 14-language mean (19.9%). The ablation studies clearly demonstrate the impact of real pause data versus segmented data, showing that the data strategy is the primary driver of performance gains. The efficiency analysis shows a 3.7x speedup for multi-question inference compared to separate passes. The integration tests with Pipecat and LiveKit frameworks provide strong evidence of practical deployability, showing reduced cut-off rates and lower latency in real-world agent setups.
The paper provides high reproducibility. The authors release the model weights, code, and framework plugins on Hugging Face. They provide detailed training hyperparameters, data mixture descriptions, and the specific attention mask implementation details. The use of open-source benchmarks (eot-bench, TurnBench) and the release of an interactive demo further enhance reproducibility and verification.
The model is evaluated primarily on English and a set of 14 languages, but the performance gap with commercial detectors in some languages (Hindi, Korean) remains. The reliance on a specific backbone (Qwen3-ASR) may limit generalizability to other architectures. The "zero-shot" performance on TurnBench is strong, but the model was trained on data that overlaps conceptually with the task, so true generalization to unseen conversational dynamics is hard to assess fully. The latency budget of 300ms is tight, and performance may degrade if hardware constraints are stricter.
This work has significant impact on the development of real-time voice agents. By providing an open-source model that matches commercial performance in turn-taking, it lowers the barrier to entry for building high-quality conversational AI. The multi-question approach is a valuable architectural pattern that can be applied to other sequential decision-making tasks in speech. The focus on real pause data sets a new standard for training data curation in this domain. VoxTurn introduces a 0.8B speech LLM that answers multiple turn-taking questions in a single forward pass using a block-isolating attention mask, achieving commercial-competitive performance on end-of-turn detection with significantly improved efficiency and multilingual robustness.
Large audio language models (LALMs) have made substantial progress in audio understanding. However, in high-throughput applications requiring frequent selection, judgment, and scoring, autoregressive text generation introduces additional computational overhead and response latency. To address this limitation, we introduce Qwen-Audio-Jev by augmenting Qwen-Audio-3.1-ASR-Flash-Next with a dedicated Jev decision head. The model unifies selection, binary judgment, and scoring through three decision primitives: CHOICE, NOUL, and SCORE. Given audio, a natural-language question, and an explicit answer specification, it directly produces a task-conditioned probability distribution or a numerical score through typed decision readouts. Evaluation across 18 labeled configurations covers all three primitives: for CHOICE, Qwen-Audio-Jev achieves accuracies of 78.70% on MMAU-mini and 84.06% on MMSU; for NOUL, it achieves an equal error rate of 6.50% on the VoxCeleb1-O speaker verification benchmark; and for SCORE, the model achieves a normalized mean absolute error of 0.1062 on NISQA overall quality assessment, with Pearson and Spearman correlations of 0.8210 and 0.8074, respectively. Compared with the original Qwen-Audio-3.1-ASR-Flash-Next, Qwen-Audio-Jev achieves an approximately 20x inference speedup. It demonstrates its potential for efficient, high-throughput audio decision-making in applications such as large-scale audio screening, speaker verification, and automated quality assessment.
Primary: Alibaba Group
All Institutions: Alibaba Group, Token Foundry
The paper introduces a unified decision head for audio LLMs that replaces autoregressive decoding with direct typed readouts for selection, judgment, and scoring, achieving significant inference speedups. While the architectural contribution is incremental, the integration of calibration-focused reinforcement learning and the comprehensive evaluation across heterogeneous audio tasks demonstrate a practical path toward efficient, high-throughput audio understanding systems.
The paper proposes a "Jev decision head" attached to the Qwen-Audio-3.1-ASR-Flash-Next backbone to replace autoregressive text generation with direct probability/score readouts for three primitives: Choice (categorical), Noul (binary), and Score (ordinal). The methodology is sound in its architectural simplicity, leveraging the existing MoE decoder's hidden states. The introduction of a Reinforcement Learning for Calibrated Decisions (RLCD) stage using Brier and ranked-probability scores is a technically interesting addition aimed at improving probability calibration, which is often neglected in standard fine-tuning. However, the core idea of adding classification heads to LLMs for specific tasks is not new, and the "Jev" terminology appears to be a proprietary or niche framework that limits the generalizability of the contribution. The training strategy of jointly optimizing heterogeneous tasks with specific loss functions (CE for Choice, BCE for Noul, and a composite loss for Score) is standard but well-executed.
The evaluation covers 18 configurations across 8 Choice, 4 Noul, and 6 Score tasks. The results are competitive, achieving 84.06% on MMSU and 6.50% EER on VoxCeleb1-O. The claim of a 20x inference speedup is significant for high-throughput applications. However, the paper lacks comparisons against other non-autoregressive audio models or specialized classifiers (e.g., ResNet-based for speaker verification, or dedicated MOS predictors like UTMOS/DNSMOS). The comparison is primarily against the autoregressive version of the same backbone, which limits the context of the performance gains. The Score results on NISQA show good correlation (0.82 Pearson) but higher MAE compared to state-of-the-art specialized MOS predictors, suggesting the unified approach may sacrifice some peak accuracy for versatility.
The paper provides architectural details (40 layers, 2048 hidden dim, 256 experts) and loss formulations. However, it does not provide code, model weights, or detailed hyperparameter settings for the RLCD stage (e.g., learning rates, number of steps, exploration temperature). The reliance on the proprietary "Qwen-Audio-3.1-ASR-Flash-Next" backbone further limits reproducibility for external researchers who may not have access to this specific checkpoint.
1. Lack of comparison with specialized non-LLM baselines for specific tasks (e.g., speaker verification, MOS prediction). 2. The "Jev" framework is not widely known, making the contribution harder to situate in the broader literature. 3. The RLCD stage is described theoretically but lacks ablation studies showing its specific contribution to calibration compared to SFT alone. 4. No streaming or real-time latency analysis beyond the theoretical speedup claim. 5. The model is evaluated on a relatively small set of benchmarks, and the "18 configurations" are not fully detailed in the provided text (tables are referenced but not fully visible in the prompt's text snippet, though the abstract summarizes key metrics).
The work addresses a practical bottleneck in deploying LALMs for high-volume tasks like screening and verification. By decoupling the decision output from text generation, it enables faster, more deterministic, and easier-to-integrate pipelines. This could facilitate the deployment of audio AI in edge devices or large-scale cloud services where latency and cost are critical. The focus on calibrated probabilities (via RLCD) is particularly valuable for safety-critical applications where uncertainty estimation is required. The paper introduces a unified decision head for audio LLMs that replaces autoregressive decoding with direct typed readouts for selection, judgment, and scoring, achieving significant inference speedups. While the architectural contribution is incremental, the integration of calibration-focused reinforcement learning and the comprehensive evaluation across heterogeneous audio tasks demonstrate a practical path toward efficient, high-throughput audio understanding systems.