Recent advancements in audio generation have been largely driven by Transformer-based diffusion models. However, these models suffer from quadratic complexity of self-attention, severely bottlenecking the long-form audio synthesis. To overcome this limitation, we propose MFlowAudio, a novel latent audio generation framework that synergizes the continuous-time dynamics of Flow Matching with a custom-designed TCMamba backbone. TCMamba uses an innovative dual-scan mechanism: a TimeMamba module to capture long-range causal dependencies with linear complexity, and a ChannelMamba module to model cross-channel correlations such as harmonic structures. Exploiting this structural foundation, we formulate the Stateful Flow Matching (SFM) paradigm. This framework inherently enables chunk-wise training and streaming generation, maintaining an $\mathcal{O}(1)$ caching complexity without incurring extra computational overhead. To enable fine-grained controllable synthesis, we devise a novel guidance mechanism that neutralizes vector field collisions precipitated by on-the-fly semantic transitions of prompts. Comprehensive empirical evaluations confirm that MFlowAudio yields generative fidelity on par with state-of-the-art baselines while establishing significant computational efficiency, achieving above $14.5$% acceleration in generation and facilitating perceptually seamless streaming synthesis with a constant latency of $1.8$ seconds. Demo: https://huggingface.co/spaces/mflowaudio/MFlowAudio
Primary: University College London
All Institutions: University College London, King's College London
The experimental evaluation is comprehensive and rigorous. The authors compare MFlowAudio against state-of-the-art full-band T2A models (AudioLDM2-48K, Stable Audio Open, TangoFlux) and an autoregressive 16kHz model (AudioGen) using a wide array of objective metrics: CLAP Score, PANNs-based Inception Score, VGGish/OpenL3-based FAD, and KL divergence. The results demonstrate that MFlowAudio achieves generative fidelity highly competitive with SOTA, even outperforming TangoFlux on several key acoustic quality metrics (FD, FADv, FADopenl3, KL divergence) with a near equivalent parameter budget. A notable strength is the detailed evaluation of computational efficiency, showing significant acceleration (81.8% speedup over TangoFlux) and the practical realization of perceptually seamless streaming generation with a constant latency of 1.8 seconds (Time-To-First-Chunk
MFlowAudio introduces a novel latent audio generation framework that combines Flow Matching with a custom-designed TCMamba backbone. The core methodological innovation lies in the TCMamba block, which addresses the limitations of Mamba for 2D spectrograms by proposing a dual-scan mechanism: TimeMamba for long-range causal dependencies along the temporal axis (linear complexity) and ChannelMamba for cross-channel correlations like harmonic structures. This is a clever adaptation of SSMs to the audio domain, specifically designed for 2D latent representations. The formulation of Stateful Flow Matching (SFM) is another significant contribution, enabling chunk-wise training and streaming generation with O(1) caching complexity. This directly tackles the memory and computational bottlenecks of traditional diffusion/flow models for long-form audio. The theoretical analysis provided in the appendix, demonstrating the equivalence of chunked and one-time integration under certain causal assumptions, provides a strong foundation for SFM. Furthermore, the State-Decoupled Conditional Guidance (SDCG) mechanism is a novel approach to handle dynamic prompt switching, aiming to neutralize vector field collisions during on-the-fly semantic transitions. This mechanism, including the Orthogonal Score Projection, is well-motivated to balance prompt adherence and acoustic continuity. The overall architecture is clearly described, leveraging standard components like VAE for audio encoding and T5 for text encoding, integrated with the TCMamba blocks. The use of FiLM and cross-attention for conditioning is standard but well-integrated.
The experimental evaluation is comprehensive and rigorous. The authors compare MFlowAudio against state-of-the-art full-band T2A models (AudioLDM2-48K, Stable Audio Open, TangoFlux) and an autoregressive 16kHz model (AudioGen) using a wide array of objective metrics: CLAP Score, PANNs-based Inception Score, VGGish/OpenL3-based FAD, and KL divergence. The results demonstrate that MFlowAudio achieves generative fidelity highly competitive with SOTA, even outperforming TangoFlux on several key acoustic quality metrics (FD, FADv, FADopenl3, KL divergence) with a near equivalent parameter budget. A notable strength is the detailed evaluation of computational efficiency, showing significant acceleration (81.8% speedup over TangoFlux) and the practical realization of perceptually seamless streaming generation with a constant latency of 1.8 seconds (Time-To-First-Chunk
Deciding which member of an animal group produced a vocalization and where that vocalization was produced are central to studying social communication but are rarely evaluated with physically verified ground truth. Existing resources either localize sounds without identifying the caller, or recognize individuals from isolated vocalizations without reliance on the 3D candidate geometry. We close this gap with V33DA, a dataset-centered multimodal benchmark for spatial vocal attribution in social zebra finches: it comprises 33,625 verified events coupling 5-channel audio with three camera views, calibrated 3D pose for every visible candidate, and per-bird radio telemetry across 10 individuals and three experiments. Caller identity is verified from an on-body accelerometer-derived vibration signal; this channel is withheld from benchmark models and used only by an on-body reference ceiling. Given a detected vocalization and the visible candidate birds, the task is to select the caller. We evaluate it under three regimes: session-disjoint testing on familiar birds from held-out recording dates, held-out-experiment transfer to a new cohort, and leave-one-individual-out evaluation on a bird whose vocalizations are held out from training. Reference methods expose a consistent gap between in-domain accuracy and transfer to unseen individuals: methods that can exploit familiar vocal identity perform well on known callers but degrade sharply on unseen ones, whereas candidate-aware methods preserving explicit spatial reasoning transfer substantially better. The on-body reference stays near-perfect across regimes, confirming the internal consistency of the accelerometer-verified labels. We release V33DA with fixed protocols and code for the evaluated methods, and V33DA++, an unevaluated companion set of verified overlapping-call events and +/-2 s context windows.
Primary: Institute of Neuroinformatics, University of Zurich and ETH Zurich
All Institutions: Institute of Neuroinformatics, University of Zurich, ETH Zurich, Swiss Data Science Center, CVLab, EPFL
[One sentence main contribution]. [Comprehensive analysis of the technical contribution, methodology, and significance to the field]. The paper introduces V33DA, a physically verified multimodal benchmark for vocal attribution in zebra finches, demonstrating that transferable attribution requires spatial reasoning rather than familiar-voice recognition, and providing a rigorous framework for evaluating multimodal models in social animal communication.
The paper proposes a rigorous methodology for constructing a physically verified benchmark for vocal attribution. The core innovation is the use of an on-body accelerometer-derived vibration signal to ground-truth caller identity, which is then withheld from the benchmark models to prevent shortcut learning. This is a sophisticated design choice that distinguishes the work from standard active speaker detection datasets that rely on perceptual annotation or close-talk references. The pipeline is detailed and auditable, involving synchronized multimodal data (5-channel audio, 3-camera video, 3D pose, radio telemetry) and a multi-stage verification process including manual review and geometric filtering. The evaluation protocol is particularly strong, defining three distinct regimes (session-disjoint, held-out-experiment, leave-one-individual-out) to explicitly separate in-domain recognition from transferable spatial reasoning.
The experimental evaluation is comprehensive, testing a wide array of baselines ranging from classical physics-based methods (SRP) to deep learning architectures (SELDnet++, EINV2, CST-former). The results clearly demonstrate the paper's central thesis: models that rely on familiar voice identity perform well in-domain but fail on unseen individuals, while spatially-aware models transfer better. The inclusion of modality intervention experiments (zeroing audio, shuffling geometry) provides strong causal evidence for what drives performance. The "on-body reference" ceiling confirms the consistency of the labels. The analysis of "close birds" and localization diagnostics adds depth to the evaluation.
Reproducibility is high. The authors release the full dataset, code for all evaluated methods, and fixed evaluation protocols. The paper provides extensive details on the data acquisition pipeline, camera calibration, and label construction in the appendices. The use of standard open-source tools (YOLOX, OC-SORT, Whisper) for preprocessing further aids reproducibility. The clear separation of the verification channel from the benchmark inputs ensures that future researchers can fairly evaluate new methods without access to the ground-truth sensor.
The primary limitation is the scale and diversity of the data. The dataset is limited to 10 zebra finches across 3 experiments in a single installation. This limits the generalizability of the findings to other species or larger, more dynamic groups. The radio telemetry features are limited due to carrier drift, which the authors acknowledge. The benchmark currently only covers single-vocalizer events, leaving multi-caller scenarios for the unevaluated V33DA++ release.
This work has significant impact on the fields of bioacoustics, multimodal learning, and behavioral analysis. By providing a physically verified benchmark, it enables the development of more robust and generalizable models for animal communication. The insights into shortcut learning and the importance of spatial reasoning are applicable to other multimodal attribution tasks, such as human active speaker detection. The release of the dataset and code will likely spur further research in this niche but important area. [One sentence main contribution]. [Comprehensive analysis of the technical contribution, methodology, and significance to the field]. The paper introduces V33DA, a physically verified multimodal benchmark for vocal attribution in zebra finches, demonstrating that transferable attribution requires spatial reasoning rather than familiar-voice recognition, and providing a rigorous framework for evaluating multimodal models in social animal communication.
We present EmbeddingGemma 2, an open multimodal embedding model which natively represents text, images, audio, video, and combinations thereof in a unified space. Our approach introduces improvements such as simplified per-layer embeddings and a multi-stage curriculum with modality and objective progression which ensure strong quality across unimodal, cross-modal, and mixed-modal evaluations. Notably, in comprehensive embedding benchmarks across all modalities, EmbeddingGemma 2 establishes a new Pareto frontier, outperforming both specialist models and models two to twelve times larger. Furthermore, our design offers flexibility for resource-constrained environments, supporting selective modality encoder loading, adjustable vision token budgets, truncatable embedding dimensions, and quantized weights. We present ablations analyzing key decisions and release EmbeddingGemma 2 to promote further research.
Primary: Google
All Institutions: Google
EmbeddingGemma 2 introduces a unified, efficient multimodal embedding model that achieves state-of-the-art performance across text, vision, and audio benchmarks while significantly reducing computational costs through architectural simplifications and modular design. The paper demonstrates that a 744M-parameter model can outperform much larger specialists by leveraging a multi-stage curriculum learning strategy that transitions from distillation to contrastive learning, effectively transferring knowledge from a larger teacher model while optimizing for the student's specific architecture and deployment constraints.
The paper proposes EmbeddingGemma 2, a 744M-parameter multimodal embedding model. The core architectural contribution is the simplification of Per-Layer Embeddings (PLE) by removing the dedicated lookup table and instead projecting the main input embeddings, which reduces memory overhead and improves multimodal performance. The training methodology is rigorous, employing a three-stage curriculum: text pre-finetuning (primarily distillation), multimodal pre-finetuning (distillation from a larger teacher model, Gemini Embedding 2), and multimodal finetuning (contrastive learning with spread-out regularization and Matryoshka Representation Learning). This progression effectively transfers knowledge from a large teacher to a smaller student while allowing the student to optimize its own representation space. The inclusion of quantization-aware training (QAT) and modular encoder loading demonstrates a strong focus on practical deployment constraints.
The evaluation is comprehensive, covering seven major benchmark suites (MTEB, MMEB-v2, MIEB-lite, MSEB, MAEB) across text, code, image, video, document, speech, and sound. The model establishes a new Pareto frontier, outperforming specialist models and general-purpose models 2-12x larger. Ablation studies are particularly strong, validating the removal of the PLE lookup table, the necessity of co-training the vision encoder, and the impact of vision token budgets. The comparison against ASR-based pipelines for audio tasks convincingly demonstrates the value of native audio understanding.
The paper provides high reproducibility for the model usage, as the checkpoints are released. It details the architecture, loss functions, and training recipe in sufficient depth. However, as is common with large-scale industrial models, the specific data mixtures and hyperparameters for the 2.9T token training run are not fully disclosed, limiting the ability to replicate the training process from scratch.
The model is still significantly larger than pure text embedding models, posing challenges for extremely low-resource edge devices despite the modular design. The paper notes that video performance degrades at higher token budgets due to attention dilution, suggesting room for improvement in long-context handling. Additionally, while it outperforms many specialists, it does not always beat the absolute best specialists in every single niche audio task (e.g., some paralinguistic clustering).
This work significantly lowers the barrier to entry for high-quality multimodal retrieval. By providing a unified, efficient, and open model, it enables developers to build complex multimodal applications (RAG, search, recommendation) without assembling complex pipelines of ASR, captioners, and separate embedding models. The focus on on-device deployment (191MB footprint for text-only) is particularly impactful for privacy-preserving applications. EmbeddingGemma 2 introduces a unified, efficient multimodal embedding model that achieves state-of-the-art performance across text, vision, and audio benchmarks while significantly reducing computational costs through architectural simplifications and modular design. The paper demonstrates that a 744M-parameter model can outperform much larger specialists by leveraging a multi-stage curriculum learning strategy that transitions from distillation to contrastive learning, effectively transferring knowledge from a larger teacher model while optimizing for the student's specific architecture and deployment constraints.
Multimodal language models continue to improve on broad music understanding benchmarks, but which musical features can they perceive and reason about? We investigate this question by introducing PianoBench, a benchmark that tasks language models with reasoning about musical performances across congruent acoustic and symbolic input representations. PianoBench is derived from our new multimodal, training-scale dataset comprising more than 20,000 hours of interleaved transcripts of music and spoken musical commentary from instructional piano videos. Open-weight audio-language models achieve at most 52.8% accuracy on PianoBench, and at best only modestly outperform their accuracy on a no-audio control. In contrast, frontier models using chain of thought achieve up to 93.2% accuracy when provided transcripts of music as text, exposing a gap between perceiving musical features from audio and reasoning about musical events when they are given explicitly. We then test whether mid-training on interleaved musical instruction can improve question answering without chain of thought, raising Qwen 3.5 9B Base's zero-shot accuracy on PianoBench with music transcripts from 42.8% to 56.0% after instruction tuning. These experiments suggest that instructional videos may represent an overlooked source of multimodal supervision for music. Available: https://gclef-cmu.org/pianobench.
Primary: Carnegie Mellon University
All Institutions: Queen Mary University of London, Carnegie Mellon University, Cornell University
The paper introduces PianoBench, a benchmark and dataset for evaluating musical perception and reasoning in language models, revealing a significant gap between audio perception and symbolic reasoning capabilities in current models. Through a rigorous data pipeline and comprehensive experiments, the authors demonstrate that while frontier models excel at reasoning over explicit note events, open-weight models struggle with audio perception, and that interleaved musical instruction can improve zero-shot reasoning without chain-of-thought.
The paper introduces a novel data pipeline to construct a large-scale (20k hours) dataset of interleaved musical commentary and symbolic transcripts from instructional piano videos. The methodology is robust, utilizing source separation, speech recognition, and MIDI transcription to create aligned multimodal data. The benchmark construction is rigorous, employing LLMs for question generation and distractor creation, followed by human verification to ensure quality. The proposed "MIDI-text" tokenization scheme is a clever contribution, allowing text-only LLMs to process symbolic music efficiently. The experimental design effectively isolates perception (audio) from reasoning (symbolic) by comparing performance across different input modalities.
The evaluation is comprehensive, testing a wide range of open-weight and proprietary models. The key findingโthat open-weight models struggle with audio perception while frontier models excel at symbolic reasoning with CoTโis well-supported by the data. The mid-training experiment on Qwen 3.5 9B Base provides strong evidence that interleaved instruction data improves zero-shot reasoning without CoT, outperforming separated data. The inclusion of a no-music control is crucial for validating the necessity of the musical input.
High. The authors release the dataset, benchmark, and code. The data pipeline is described in detail, including the specific models used for transcription and separation. The hyperparameters for mid-training are provided. The use of specific commercial APIs (GPT 5.6, Gemini 3.1) limits full reproducibility for those specific results, but the open-weight model results are reproducible.
The benchmark is limited to piano music, which may not generalize to other instruments or genres. The reliance on YouTube instructional videos may introduce biases in repertoire and teaching style. The "MIDI-text" representation, while efficient, is a lossy compression of the original audio and may not capture timbral or expressive nuances. The small size of the benchmark (250 questions) limits statistical power for fine-grained analysis.
This work highlights a critical gap in current audio-language models: the ability to reason about music when given symbolic representations versus perceiving it from audio. It provides a valuable resource (PianoTalk dataset) for training models on fine-grained musical understanding. The findings suggest that current open-weight models need better supervision for audio perception, while frontier models' reasoning capabilities can be leveraged to generate better training data. This could drive future research in multimodal music understanding and the development of more perceptually grounded audio models. The paper introduces PianoBench, a benchmark and dataset for evaluating musical perception and reasoning in language models, revealing a significant gap between audio perception and symbolic reasoning capabilities in current models. Through a rigorous data pipeline and comprehensive experiments, the authors demonstrate that while frontier models excel at reasoning over explicit note events, open-weight models struggle with audio perception, and that interleaved musical instruction can improve zero-shot reasoning without chain-of-thought.
Spatial audio conveys not only what sounds are present, but where they originate and how they interact with the environment. Recent audio-language models increasingly understand and reason about such scenes, yet spatial audio understanding, generation, and editing remain largely separate. We present Spatial Transfusion, a unified first-order ambisonics (FOA) framework that extends a pretrained spatial audio understanding model to generation and editing through flow matching over VAE latents. We formulate editing as a structured scene transformation over a taxonomy of atomic edits. Keeping the pretrained backbone frozen, we introduce modality-based LoRA routing and a curriculum that progressively adds generation, editing, and natural-language interaction. Experiments show that Spatial Transfusion largely preserves spatial understanding while outperforming full fine-tuning in generation and editing. The resulting model supports controllable FOA generation and editing through both structured and natural-language interfaces. Audio samples are available at https://microsoft.github.io/spatial-transfusion/
Primary: Carnegie Mellon University
All Institutions: Carnegie Mellon University, Microsoft Corporation
Spatial Transfusion introduces a unified framework for FOA audio understanding, generation, and editing by extending a pretrained spatial audio language model with flow matching and modality-based LoRA routing. The paper demonstrates that this approach preserves spatial understanding while enabling high-quality, controllable generation and editing, outperforming full fine-tuning baselines and establishing a new paradigm for spatial audio manipulation.
The paper proposes Spatial Transfusion, a unified framework for First-Order Ambisonics (FOA) audio understanding, generation, and editing. The core methodological contribution is the adaptation of a pretrained spatial audio language model (Sci-Phi) using a "Transfusion" paradigm, which combines autoregressive text modeling with flow-matching for continuous audio latents. A key technical innovation is the use of modality-based LoRA routing, where separate LoRA adapters are applied to audio and text tokens, allowing the frozen backbone to learn generative capabilities without catastrophic forgetting of understanding tasks. The authors define a structured taxonomy of atomic edits (semantic, acoustic, spatial, temporal) and employ a four-stage curriculum to progressively introduce generation and editing capabilities. The approach effectively leverages the structured nature of spatial scenes to enable precise control over generation and editing, distinguishing it from generic audio generation models.
The experimental setup is rigorous, utilizing a held-out set from the FOA-MEIR dataset with real room impulse responses. The evaluation metrics are comprehensive, covering spatial fidelity (FAD-PSELD, FOA covariance), semantic alignment (CLAP), and editing-specific metrics (EditFLAM). The results demonstrate that the proposed LoRA routing strategy outperforms full fine-tuning in generation and editing while preserving understanding capabilities. Ablation studies on routing strategies (modality-based vs. task-based vs. dense) provide strong evidence for the effectiveness of the chosen approach. The inclusion of natural language instruction tuning (Stage 4) further validates the flexibility of the framework.
The paper provides detailed descriptions of the architecture, training curriculum, and hyperparameters. However, specific details on the FOA-VAE implementation and the exact simulation pipeline parameters are referenced to prior work (Sci-Phi, FOA-VAE), which may limit immediate reproducibility without access to those codebases. The availability of a demo page with audio samples is a positive indicator, but a full code release is not explicitly mentioned in the text provided.
The model is trained primarily on simulated data, which may limit its generalization to real-world noisy environments, although the evaluation uses real RIRs. The framework is specific to FOA, and extension to higher-order ambisonics or binaural audio is not explored. The editing operations are limited to the defined atomic taxonomy, and complex, multi-step edits may not be supported. The reliance on a specific pretrained backbone (Phi-4-MM) may constrain the model's scalability.
This work bridges the gap between spatial audio understanding and generation, offering a unified interface for immersive audio creation. The structured editing capabilities have potential applications in virtual reality, gaming, and film production, where precise control over soundscapes is crucial. The framework's ability to handle both structured JSON and natural language inputs makes it accessible to both technical and non-technical users. Spatial Transfusion introduces a unified framework for FOA audio understanding, generation, and editing by extending a pretrained spatial audio language model with flow matching and modality-based LoRA routing. The paper demonstrates that this approach preserves spatial understanding while enabling high-quality, controllable generation and editing, outperforming full fine-tuning baselines and establishing a new paradigm for spatial audio manipulation.
Generative spoken language models (SLMs) offer a route to learning linguistic structure purely from speech, yet they lag far behind text-based models in both capability and data scaling efficiency. We identify two bottlenecks underlying this gap: the limited linguistic quality of existing speech units and the reliance on a single stream of discrete units to represent speech. Multiple codebooks are widely used to encode acoustic features but rarely to represent speech semantics. We present GSLM2, which learns and models complementary properties of speech at multiple levels of abstraction. Its encoder, SpidR 2.0, uses self-supervised learning with differentiable prediction heads to induce high-quality discrete units at multiple levels. A multi-stream SLM then jointly models these units using a shared temporal backbone, stream-specific predictors, and an autoregressive head that captures cross-stream dependencies. Across training data sizes and model scales, up to 6M hours and 6.5B parameters, GSLM2 consistently improves linguistic competence while scaling more efficiently with data. As a result, it substantially narrows the gap between spoken and text-based LMs; across read and conversational speech domains, we shrink the gap by about two orders of magnitude on sWuggy, sBLIMP, and tSC, while also outperforming competitive speech-text systems on all three tasks. We open-source the training code and model checkpoints at https://github.com/facebookresearch/gslm2.
Primary: Meta Superintelligence Labs
All Institutions: Meta Superintelligence Labs, Universite Paris Dauphine-PSL, AMI Labs, Doctolib, Ecole des Hautes Etudes en Sciences Sociales
GSLM2 introduces a multi-stream speech representation learning framework that significantly improves the data efficiency and linguistic competence of generative spoken language models. By jointly modeling discrete units from multiple encoder layers using a shared backbone and cross-stream generative heads, the method effectively captures complementary phonetic and semantic information, substantially narrowing the performance gap with text-based language models and outperforming existing speech-text systems on key linguistic benchmarks.
The paper proposes GSLM2, a framework addressing the data efficiency gap between spoken and text language models. The core methodological contribution is twofold: (1) SpidR 2.0, an encoder pretraining method that uses differentiable prediction heads to induce high-quality discrete units at multiple levels of abstraction (layers) simultaneously, replacing post-hoc k-means with joint optimization; and (2) a multi-stream SLM architecture that jointly models these unit streams using a shared temporal backbone, stream-specific predictors, and an autoregressive generative head to capture cross-stream dependencies. The use of entropy regularization to stabilize code distributions and the specific training objective combining single-stream, self-distillation, and multi-stream generative losses are technically sound and well-motivated.
The experiments are extensive, scaling up to 6M hours of data and 6.5B parameters. The paper demonstrates that GSLM2 consistently improves linguistic competence and scales more efficiently with data compared to single-stream baselines. It claims to shrink the gap with text-based LMs by about two orders of magnitude on key benchmarks (sWuggy, sBLIMP, tSC) and outperforms competitive speech-text systems. The comparison against both speech-only and speech-text baselines provides a strong evaluation of the textless approach.
The authors state that they open-source the training code and model checkpoints at the provided GitHub URL. The paper provides detailed mathematical formulations for the encoder pretraining and SLM training objectives, which aids in understanding the implementation. However, specific hyperparameters for the 6M hour runs may be complex to replicate without the full codebase, but the release of checkpoints is a significant positive factor.
The paper focuses primarily on English speech (Libri-Light/LibriSpeech). While it claims generalizability, the evaluation is limited to specific linguistic tasks (sWuggy, sBLIMP, tSC) which may not fully capture the breadth of spoken language understanding. The computational cost of training multi-stream models at scale is high, though the data efficiency gains are highlighted. The paper is an arXiv preprint, so peer review status is pending.
This work is significant for the field of speech AI as it provides a scalable, textless approach to learning linguistic structure. By narrowing the gap with text LMs, it enables more robust spoken dialogue systems that do not rely on transcriptions, which is crucial for low-resource languages or real-time applications. The open-sourcing of the code and checkpoints will likely accelerate research in multi-stream speech representation learning. GSLM2 introduces a multi-stream speech representation learning framework that significantly improves the data efficiency and linguistic competence of generative spoken language models. By jointly modeling discrete units from multiple encoder layers using a shared backbone and cross-stream generative heads, the method effectively captures complementary phonetic and semantic information, substantially narrowing the performance gap with text-based language models and outperforming existing speech-text systems on key linguistic benchmarks.
Speech-to-speech translation (S2ST) increasingly preserves what is said and how it sounds, but not necessarily when speech and silence occur. For dubbing and other temporally grounded applications, matching total duration is not enough: speech and pauses must also occur at corresponding moments. We formulate timeline-preserving S2ST, which aims to preserve the source endpoint and salient internal pauses on a shared absolute timeline. We present TransDubber, an end-to-end autoregressive system that learns this temporal structure directly from aligned target waveforms, without inference-time frame-level timing control or post-hoc temporal correction. To provide this supervision, we construct 642K exactly duration-matched English-Chinese speech pairs by translating and synthesizing target speech on source-defined temporal scaffolds. Training combines a progressive supervised curriculum with verifier-guided reinforcement learning to balance translation quality and temporal alignment. On a 4,353-utterance test set spanning read and spontaneous speech, TransDubber increases the proportion of translations within 2% of the source duration from 25.6% to 61.5% and pause-alignment F1 from 0.118 to 0.550 compared with UniSS, while maintaining competitive translation quality. Our analysis shows that aligned supervision establishes timeline following, while reinforcement learning refines temporal precision and partially offsets the translation-quality cost of strict synchronization. Together, these results demonstrate the effectiveness of our data and training framework for timeline-preserving S2ST. Audio samples, code, datasets, and pretrained models are available at https://zhishengzheng.com/transdubber/.
Primary: The University of Texas at Austin
All Institutions: The University of Texas at Austin
The paper introduces a rigorous framework for timeline-preserving speech-to-speech translation, combining a scalable synthetic data pipeline with component-specific reinforcement learning to achieve state-of-the-art temporal alignment without compromising translation quality.
The paper proposes a novel formulation of "timeline-preserving" Speech-to-Speech Translation (S2ST), distinguishing it from prior work that focuses solely on total duration matching or speaker identity. The core methodological contribution is a data construction pipeline that creates 642K exactly duration-matched English-Chinese speech pairs by synthesizing target speech on "source-defined temporal scaffolds." This involves extracting speech/pause regions from source audio, translating segments under local duration budgets, and rendering them back onto the source timeline. The training strategy is a progressive curriculum: (1) pause-aware ASR/translation, (2) natural-pace speech generation, (3) timeline-preserving supervised fine-tuning, and (4) verifier-guided reinforcement learning (GSPO). The RL stage is particularly notable for its component-specific credit assignment, where translation quality rewards are applied to text tokens, while duration and pause alignment rewards are applied to speech-codec tokens, avoiding the dilution of signals that occurs in scalar reward aggregation.
The experiments are rigorous, utilizing a 4,353-utterance test set spanning read (CoVoST2) and spontaneous (GigaSpeech) speech. The system significantly outperforms strong baselines (UniSS, Seamless, Cascaded) in temporal metrics: increasing the proportion of translations within 2% of source duration from 25.6% to 61.5% and pause-alignment F1 from 0.118 to 0.550. Crucially, this is achieved without sacrificing translation quality (COMET 0.788 vs 0.784 for UniSS) or speaker similarity. Ablation studies effectively isolate the contributions of pause markers, the natural-pace warm-up stage, and individual reward components, demonstrating that the RL stage specifically refines temporal precision while partially offsetting the quality cost of strict synchronization.
The paper provides high reproducibility. It details the exact data construction steps, including VAD thresholds, duration budget calculations, and synthesis parameters. The authors commit to releasing code, datasets, and pretrained models. The use of open-source components (Qwen3-ASR, VoxCPM2, Hy-MT2) for the pipeline further aids reproducibility, although the specific "source-canvas rendering" logic is a custom contribution that requires careful implementation.
The primary limitation is the reliance on synthetic data for supervision. While the pipeline is scalable, the quality of the synthesized "timeline-preserving" targets depends heavily on the underlying TTS model (VoxCPM2) and the translation model's ability to adhere to strict character budgets. The evaluation is limited to English-Chinese; generalization to other language pairs with different prosodic structures is not tested. Additionally, the "timeline preservation" constraint may not always be desirable in all dubbing contexts, as noted in the related work, though the paper frames this as an application-specific choice.
This work has significant implications for the dubbing industry and real-time speech translation applications where temporal synchronization with visual cues or other audio tracks is critical. By demonstrating that autoregressive models can learn complex temporal structures directly from aligned waveforms without explicit frame-level control at inference, it opens the door to more natural and synchronized multilingual audio generation. The component-specific RL approach could be adapted to other multimodal generation tasks requiring precise alignment between different modalities. The paper introduces a rigorous framework for timeline-preserving speech-to-speech translation, combining a scalable synthetic data pipeline with component-specific reinforcement learning to achieve state-of-the-art temporal alignment without compromising translation quality.
Recent advancements in audio generation have been largely driven by Transformer-based diffusion models. However, these models suffer from quadratic complexity of self-attention, severely bottlenecking the long-form audio synthesis. To overcome this limitation, we propose MFlowAudio, a novel latent audio generation framework that synergizes the continuous-time dynamics of Flow Matching with a custom-designed TCMamba backbone. TCMamba uses an innovative dual-scan mechanism: a TimeMamba module to capture long-range causal dependencies with linear complexity, and a ChannelMamba module to model cross-channel correlations such as harmonic structures. Exploiting this structural foundation, we formulate the Stateful Flow Matching (SFM) paradigm. This framework inherently enables chunk-wise training and streaming generation, maintaining an $\mathcal{O}(1)$ caching complexity without incurring extra computational overhead. To enable fine-grained controllable synthesis, we devise a novel guidance mechanism that neutralizes vector field collisions precipitated by on-the-fly semantic transitions of prompts. Comprehensive empirical evaluations confirm that MFlowAudio yields generative fidelity on par with state-of-the-art baselines while establishing significant computational efficiency, achieving above $14.5$% acceleration in generation and facilitating perceptually seamless streaming synthesis with a constant latency of $1.8$ seconds. Demo: https://huggingface.co/spaces/mflowaudio/MFlowAudio
Primary: University College London
All Institutions: University College London, King's College London
The experimental evaluation is comprehensive and rigorous. The authors compare MFlowAudio against state-of-the-art full-band T2A models (AudioLDM2-48K, Stable Audio Open, TangoFlux) and an autoregressive 16kHz model (AudioGen) using a wide array of objective metrics: CLAP Score, PANNs-based Inception Score, VGGish/OpenL3-based FAD, and KL divergence. The results demonstrate that MFlowAudio achieves generative fidelity highly competitive with SOTA, even outperforming TangoFlux on several key acoustic quality metrics (FD, FADv, FADopenl3, KL divergence) with a near equivalent parameter budget. A notable strength is the detailed evaluation of computational efficiency, showing significant acceleration (81.8% speedup over TangoFlux) and the practical realization of perceptually seamless streaming generation with a constant latency of 1.8 seconds (Time-To-First-Chunk
MFlowAudio introduces a novel latent audio generation framework that combines Flow Matching with a custom-designed TCMamba backbone. The core methodological innovation lies in the TCMamba block, which addresses the limitations of Mamba for 2D spectrograms by proposing a dual-scan mechanism: TimeMamba for long-range causal dependencies along the temporal axis (linear complexity) and ChannelMamba for cross-channel correlations like harmonic structures. This is a clever adaptation of SSMs to the audio domain, specifically designed for 2D latent representations. The formulation of Stateful Flow Matching (SFM) is another significant contribution, enabling chunk-wise training and streaming generation with O(1) caching complexity. This directly tackles the memory and computational bottlenecks of traditional diffusion/flow models for long-form audio. The theoretical analysis provided in the appendix, demonstrating the equivalence of chunked and one-time integration under certain causal assumptions, provides a strong foundation for SFM. Furthermore, the State-Decoupled Conditional Guidance (SDCG) mechanism is a novel approach to handle dynamic prompt switching, aiming to neutralize vector field collisions during on-the-fly semantic transitions. This mechanism, including the Orthogonal Score Projection, is well-motivated to balance prompt adherence and acoustic continuity. The overall architecture is clearly described, leveraging standard components like VAE for audio encoding and T5 for text encoding, integrated with the TCMamba blocks. The use of FiLM and cross-attention for conditioning is standard but well-integrated.
The experimental evaluation is comprehensive and rigorous. The authors compare MFlowAudio against state-of-the-art full-band T2A models (AudioLDM2-48K, Stable Audio Open, TangoFlux) and an autoregressive 16kHz model (AudioGen) using a wide array of objective metrics: CLAP Score, PANNs-based Inception Score, VGGish/OpenL3-based FAD, and KL divergence. The results demonstrate that MFlowAudio achieves generative fidelity highly competitive with SOTA, even outperforming TangoFlux on several key acoustic quality metrics (FD, FADv, FADopenl3, KL divergence) with a near equivalent parameter budget. A notable strength is the detailed evaluation of computational efficiency, showing significant acceleration (81.8% speedup over TangoFlux) and the practical realization of perceptually seamless streaming generation with a constant latency of 1.8 seconds (Time-To-First-Chunk
Recent progress in visual scene generation has extended beyond static 3D scenes toward the more challenging problem of generating dynamic 4D scenes from minimal input. However, sound, despite being intrinsically coupled with dynamic objects and events, remains largely unmodeled in prior work. We introduce Video-to-4DAVScene, a new task that turns a video into an explorable, time-varying 4D audio-visual scene, and present the SonoWorld4D framework for this task. Given a monocular video, our method generates a 360 static background, reconstructs explicit motion trajectories of dynamic, sound-emitting foreground objects, and synthesizes a physically-grounded sound field consistent across multiple sources through dual-space audio diffusion synchronization. The resulting 4D scene supports real-time novel-view visual and spatial audio rendering. Evaluation on real-world dynamic scenes and a user study demonstrates SonoWorld4D generates 4D audio-visual scenes with perceptually realistic, physically grounded spatial audio consistent with scene geometry and visual content.
Primary: University of Maryland, College Park
All Institutions: University of Maryland, College Park
The paper introduces a novel task of Video-to-4DAVScene and a framework that effectively synchronizes visual dynamics with spatial audio using dual-space diffusion. The technical approach is rigorous, combining explicit 4D geometry with physically grounded acoustic rendering, and the introduction of a new dataset and evaluation protocol significantly advances the field of dynamic audio-visual scene generation.
The paper proposes SonoWorld4D, a framework for generating explorable 4D audio-visual scenes from a single monocular video. The methodology is technically sophisticated, relying on a decomposed 4D visual representation (static background Gaussians + dynamic foreground objects with explicit trajectories) and a novel "dual-space diffusion synchronization" mechanism. This synchronization couples camera-space audio generation with object-space audio generation via forward and inverse acoustic rendering operators during the diffusion sampling process. The use of flow-matching and the specific bidirectional consistency constraints (rendering object sounds to camera space and separating/inverting camera sounds to object space) represent a strong technical contribution to the field of spatial audio generation. The integration of VLM-derived prompts for acoustic context is also a relevant modern addition.
The authors introduce a new dataset (SonoScene4D) with 54 clips of dynamic scenarios featuring moving sound sources, recorded with synchronized FOA audio at novel listener poses. This is a valuable contribution to the field, as existing datasets often lack dynamic spatial audio ground truth. The evaluation includes objective metrics (DoA errors, CC, AUC, CLAP) and a user study with 50 participants. The results show significant improvements over baselines like Sonic4D and OmniAudio in both spatial accuracy and semantic consistency. The qualitative analysis of Doppler effects and novel listener adaptation is convincing.
The paper provides detailed descriptions of the pipeline, including the specific models used for depth estimation, segmentation, and Gaussian generation. However, the reliance on several proprietary or very recent models (e.g., HYWorld 2.0, Marble, SAM3, Kimi VLM) may limit immediate reproducibility for the community. The code is not explicitly linked in the provided text (only a demo page), which is a minor drawback for open science, though the demo page likely hosts assets.
The method assumes outdoor scenes with limited reverberation, omitting reflected paths and complex room acoustics, which limits its applicability to indoor environments. The pipeline is complex and relies on a chain of foundation models, which may introduce error propagation. The dataset size (54 clips) is relatively small for robust statistical validation, though it is a new task. The computational cost of the dual-space diffusion synchronization and 4D Gaussian rendering is likely high, though real-time rendering is claimed for the final scene.
This work bridges the gap between 4D visual generation and spatial audio synthesis, enabling the creation of immersive, interactive worlds from minimal input. It has potential applications in virtual reality, game asset generation, and film pre-visualization. The introduction of a benchmark for dynamic spatial audio generation is particularly impactful for the research community. The paper introduces a novel task of Video-to-4DAVScene and a framework that effectively synchronizes visual dynamics with spatial audio using dual-space diffusion. The technical approach is rigorous, combining explicit 4D geometry with physically grounded acoustic rendering, and the introduction of a new dataset and evaluation protocol significantly advances the field of dynamic audio-visual scene generation.
A capable voice assistant must do more than recognize words; it must understand requests, reason over context, sustain conversations, and act through tools. We present Apriel Voice 27B Preview, an open speech-language model that unifies these capabilities within a shared reasoning backbone while retaining text and image input modalities. Built on the post-trained Gemma 4 26B-A4B checkpoint, Apriel Voice directly processes speech and generates both responses and tool calls without requiring a separate dialogue model. We introduce a two-stage training procedure to adapt the model to speech. The first stage establishes the speech interface, while the second jointly adapts the projector and backbone through supervised fine-tuning, text replay, and source-dependent cross-entropy and forward Kullback--Leibler divergence supervision. Automatically constructed speech-conditioned dialogues provide conversational training data, while local teacher serving, overlapping teacher and student computation, and fused full-vocabulary losses enable efficient online distillation. In thinking mode, Apriel Voice achieves a score of 98.90 on Big Bench Audio, 81.13 on Speech-BFCL, and a task success rate of 65.51% on $ฯ$-Voice, with a mean latency of 1.91 seconds. It also achieves 51.27% English EVA accuracy and 74.13% experience pass@1. Finally, we deploy Apriel Voice 27B Preview on a custom serving platform to support interactive voice interactions.
Primary: ServiceNow
All Institutions: ServiceNow
Apriel Voice 27B Preview is an open speech-language model that unifies speech understanding, reasoning, and tool use within a shared backbone. The paper presents a rigorous two-stage adaptation recipe combining projector training with joint fine-tuning using source-dependent CE/FKL supervision, demonstrating strong performance in agentic speech tasks while retaining text and image capabilities.
The paper proposes a two-stage training procedure to adapt a text-and-image LLM (Gemma 4 26B-A4B) into a speech-language model. Stage 1 freezes the backbone and trains a speech projector (Whisper encoder + MLP) on transcription, translation, and continuation tasks. Stage 2 jointly fine-tunes the projector and backbone using a mixture of speech tasks, synthetic speech-conditioned dialogues, tool interactions, and text replay. A key technical contribution is the source-dependent loss function combining Cross-Entropy (CE) for task targets and Forward Kullback-Leibler (FKL) divergence from a frozen teacher model to preserve original capabilities. The system also details an efficient training infrastructure using local teacher serving, overlapping computation, and fused full-vocabulary loss kernels. While the individual components (projector, distillation, SFT) are established, the specific combination for agentic speech tasks and the engineering optimizations for online distillation are well-executed.
The evaluation is comprehensive, covering foundational speech metrics (WER, BLEU), reasoning (Big Bench Audio, GSM8K), instruction following (IFEval), and agentic task completion (EVA, ฯ-Voice, EnterpriseOps-Gym). The model demonstrates strong performance, particularly in reasoning (98.90 on BBA) and tool use (81.13 on BFCL), outperforming other open-source 2-stage models. However, the reliance on synthetic data for conversation training and the lack of ablation studies isolating the impact of specific training components (e.g., FKL vs. CE, specific data mixes) limit the depth of the empirical analysis. The comparison with proprietary systems is competitive but relies on different evaluation harnesses and simulators, making direct comparison difficult.
The paper provides detailed architectural specifications (projector dimensions, parameter counts) and training hyperparameters (learning rates, batch sizes, GPU configuration). The release of an interactive demo and the use of open-source base models (Gemma, Whisper) enhance reproducibility. However, the synthetic data generation pipeline and specific internal datasets are not fully released, and the "thinking mode" implementation details are not fully specified, which may hinder exact reproduction of the reported reasoning scores.
The model is primarily optimized for English, with limited multilingual speech understanding. The evaluation lacks paired significance tests against the base model, making it hard to quantify the exact cost of adaptation on text/image capabilities. The synthetic data pipeline may propagate errors from the TTS or LLM generators. The paper acknowledges that frozen visual components do not guarantee unchanged image reasoning, yet the drop in some vision benchmarks suggests potential degradation.
This work contributes to the development of open-source speech-language models capable of agentic tasks, which is a critical area for enterprise and consumer applications. By demonstrating that a single backbone can handle speech, text, images, and tool calls, it supports the trend toward unified multimodal agents. The efficient training techniques for online distillation are valuable for the broader community adapting large models to new modalities. Apriel Voice 27B Preview is an open speech-language model that unifies speech understanding, reasoning, and tool use within a shared backbone. The paper presents a rigorous two-stage adaptation recipe combining projector training with joint fine-tuning using source-dependent CE/FKL supervision, demonstrating strong performance in agentic speech tasks while retaining text and image capabilities.
Deciding which member of an animal group produced a vocalization and where that vocalization was produced are central to studying social communication but are rarely evaluated with physically verified ground truth. Existing resources either localize sounds without identifying the caller, or recognize individuals from isolated vocalizations without reliance on the 3D candidate geometry. We close this gap with V33DA, a dataset-centered multimodal benchmark for spatial vocal attribution in social zebra finches: it comprises 33,625 verified events coupling 5-channel audio with three camera views, calibrated 3D pose for every visible candidate, and per-bird radio telemetry across 10 individuals and three experiments. Caller identity is verified from an on-body accelerometer-derived vibration signal; this channel is withheld from benchmark models and used only by an on-body reference ceiling. Given a detected vocalization and the visible candidate birds, the task is to select the caller. We evaluate it under three regimes: session-disjoint testing on familiar birds from held-out recording dates, held-out-experiment transfer to a new cohort, and leave-one-individual-out evaluation on a bird whose vocalizations are held out from training. Reference methods expose a consistent gap between in-domain accuracy and transfer to unseen individuals: methods that can exploit familiar vocal identity perform well on known callers but degrade sharply on unseen ones, whereas candidate-aware methods preserving explicit spatial reasoning transfer substantially better. The on-body reference stays near-perfect across regimes, confirming the internal consistency of the accelerometer-verified labels. We release V33DA with fixed protocols and code for the evaluated methods, and V33DA++, an unevaluated companion set of verified overlapping-call events and +/-2 s context windows.
Primary: Institute of Neuroinformatics, University of Zurich and ETH Zurich
All Institutions: Institute of Neuroinformatics, University of Zurich, ETH Zurich, Swiss Data Science Center, CVLab, EPFL
[One sentence main contribution]. [Comprehensive analysis of the technical contribution, methodology, and significance to the field]. The paper introduces V33DA, a physically verified multimodal benchmark for vocal attribution in zebra finches, demonstrating that transferable attribution requires spatial reasoning rather than familiar-voice recognition, and providing a rigorous framework for evaluating multimodal models in social animal communication.
The paper proposes a rigorous methodology for constructing a physically verified benchmark for vocal attribution. The core innovation is the use of an on-body accelerometer-derived vibration signal to ground-truth caller identity, which is then withheld from the benchmark models to prevent shortcut learning. This is a sophisticated design choice that distinguishes the work from standard active speaker detection datasets that rely on perceptual annotation or close-talk references. The pipeline is detailed and auditable, involving synchronized multimodal data (5-channel audio, 3-camera video, 3D pose, radio telemetry) and a multi-stage verification process including manual review and geometric filtering. The evaluation protocol is particularly strong, defining three distinct regimes (session-disjoint, held-out-experiment, leave-one-individual-out) to explicitly separate in-domain recognition from transferable spatial reasoning.
The experimental evaluation is comprehensive, testing a wide array of baselines ranging from classical physics-based methods (SRP) to deep learning architectures (SELDnet++, EINV2, CST-former). The results clearly demonstrate the paper's central thesis: models that rely on familiar voice identity perform well in-domain but fail on unseen individuals, while spatially-aware models transfer better. The inclusion of modality intervention experiments (zeroing audio, shuffling geometry) provides strong causal evidence for what drives performance. The "on-body reference" ceiling confirms the consistency of the labels. The analysis of "close birds" and localization diagnostics adds depth to the evaluation.
Reproducibility is high. The authors release the full dataset, code for all evaluated methods, and fixed evaluation protocols. The paper provides extensive details on the data acquisition pipeline, camera calibration, and label construction in the appendices. The use of standard open-source tools (YOLOX, OC-SORT, Whisper) for preprocessing further aids reproducibility. The clear separation of the verification channel from the benchmark inputs ensures that future researchers can fairly evaluate new methods without access to the ground-truth sensor.
The primary limitation is the scale and diversity of the data. The dataset is limited to 10 zebra finches across 3 experiments in a single installation. This limits the generalizability of the findings to other species or larger, more dynamic groups. The radio telemetry features are limited due to carrier drift, which the authors acknowledge. The benchmark currently only covers single-vocalizer events, leaving multi-caller scenarios for the unevaluated V33DA++ release.
This work has significant impact on the fields of bioacoustics, multimodal learning, and behavioral analysis. By providing a physically verified benchmark, it enables the development of more robust and generalizable models for animal communication. The insights into shortcut learning and the importance of spatial reasoning are applicable to other multimodal attribution tasks, such as human active speaker detection. The release of the dataset and code will likely spur further research in this niche but important area. [One sentence main contribution]. [Comprehensive analysis of the technical contribution, methodology, and significance to the field]. The paper introduces V33DA, a physically verified multimodal benchmark for vocal attribution in zebra finches, demonstrating that transferable attribution requires spatial reasoning rather than familiar-voice recognition, and providing a rigorous framework for evaluating multimodal models in social animal communication.
We present EmbeddingGemma 2, an open multimodal embedding model which natively represents text, images, audio, video, and combinations thereof in a unified space. Our approach introduces improvements such as simplified per-layer embeddings and a multi-stage curriculum with modality and objective progression which ensure strong quality across unimodal, cross-modal, and mixed-modal evaluations. Notably, in comprehensive embedding benchmarks across all modalities, EmbeddingGemma 2 establishes a new Pareto frontier, outperforming both specialist models and models two to twelve times larger. Furthermore, our design offers flexibility for resource-constrained environments, supporting selective modality encoder loading, adjustable vision token budgets, truncatable embedding dimensions, and quantized weights. We present ablations analyzing key decisions and release EmbeddingGemma 2 to promote further research.
Primary: Google
All Institutions: Google
EmbeddingGemma 2 introduces a unified, efficient multimodal embedding model that achieves state-of-the-art performance across text, vision, and audio benchmarks while significantly reducing computational costs through architectural simplifications and modular design. The paper demonstrates that a 744M-parameter model can outperform much larger specialists by leveraging a multi-stage curriculum learning strategy that transitions from distillation to contrastive learning, effectively transferring knowledge from a larger teacher model while optimizing for the student's specific architecture and deployment constraints.
The paper proposes EmbeddingGemma 2, a 744M-parameter multimodal embedding model. The core architectural contribution is the simplification of Per-Layer Embeddings (PLE) by removing the dedicated lookup table and instead projecting the main input embeddings, which reduces memory overhead and improves multimodal performance. The training methodology is rigorous, employing a three-stage curriculum: text pre-finetuning (primarily distillation), multimodal pre-finetuning (distillation from a larger teacher model, Gemini Embedding 2), and multimodal finetuning (contrastive learning with spread-out regularization and Matryoshka Representation Learning). This progression effectively transfers knowledge from a large teacher to a smaller student while allowing the student to optimize its own representation space. The inclusion of quantization-aware training (QAT) and modular encoder loading demonstrates a strong focus on practical deployment constraints.
The evaluation is comprehensive, covering seven major benchmark suites (MTEB, MMEB-v2, MIEB-lite, MSEB, MAEB) across text, code, image, video, document, speech, and sound. The model establishes a new Pareto frontier, outperforming specialist models and general-purpose models 2-12x larger. Ablation studies are particularly strong, validating the removal of the PLE lookup table, the necessity of co-training the vision encoder, and the impact of vision token budgets. The comparison against ASR-based pipelines for audio tasks convincingly demonstrates the value of native audio understanding.
The paper provides high reproducibility for the model usage, as the checkpoints are released. It details the architecture, loss functions, and training recipe in sufficient depth. However, as is common with large-scale industrial models, the specific data mixtures and hyperparameters for the 2.9T token training run are not fully disclosed, limiting the ability to replicate the training process from scratch.
The model is still significantly larger than pure text embedding models, posing challenges for extremely low-resource edge devices despite the modular design. The paper notes that video performance degrades at higher token budgets due to attention dilution, suggesting room for improvement in long-context handling. Additionally, while it outperforms many specialists, it does not always beat the absolute best specialists in every single niche audio task (e.g., some paralinguistic clustering).
This work significantly lowers the barrier to entry for high-quality multimodal retrieval. By providing a unified, efficient, and open model, it enables developers to build complex multimodal applications (RAG, search, recommendation) without assembling complex pipelines of ASR, captioners, and separate embedding models. The focus on on-device deployment (191MB footprint for text-only) is particularly impactful for privacy-preserving applications. EmbeddingGemma 2 introduces a unified, efficient multimodal embedding model that achieves state-of-the-art performance across text, vision, and audio benchmarks while significantly reducing computational costs through architectural simplifications and modular design. The paper demonstrates that a 744M-parameter model can outperform much larger specialists by leveraging a multi-stage curriculum learning strategy that transitions from distillation to contrastive learning, effectively transferring knowledge from a larger teacher model while optimizing for the student's specific architecture and deployment constraints.
Real-world audio applications require precise, queryable representations grounded in physical time. Yet modern Large Audio-Language Models (LALMs) increasingly favor conversational generation, producing free-form descriptions that conflate acoustic perception with higher-level reasoning and lack precise temporal grounding. We introduce TAC2, a foundation model that casts timestamped audio captioning as composable functions to construct granular, speaker-attributed acoustic timelines directly from in-the-wild audio. To avoid prohibitively expensive manual annotation, we programmatically fuse complementary predictions from specialized weak teachers through time-registered IoU matching. We model the resulting multi-granularity schema of timestamped events, speaker turns, and fine-grained speech attributes using an Instruction-Input-Output (IIO) completion format, in which each task is a function of the audio, an instruction, and an optional input. A single call decodes the full description autonomously, while composing calls lets earlier outputs or external priors, such as a reference transcript, condition later ones. We further extend TAC2 with expressive speech attributes to obtain TAC2-S, using parameter-efficient specialists to elicit latent paralinguistic capabilities from the pre-trained backbone and distill them into a unified model. We validate the resulting representation in two complementary settings. Paired with frozen text-only LLMs, it yields strong spoken audio understanding across diverse reasoning backends, demonstrating the utility of decoupling fine-grained acoustic perception from downstream reasoning. Used as paired supervision for text-to-audio generation, it supports semantically and temporally grounded reconstruction, together with explicit zero-shot control over event timing, speaker attributes, and word-level emphasis.
Primary: Adobe Research
All Institutions: University of Surrey, Adobe Research
TAC2 introduces a scalable framework for timestamped audio captioning by fusing weak teacher predictions and eliciting latent paralinguistic capabilities, resulting in a composable, temporally grounded representation that enhances both audio understanding and controllable generation.
The paper proposes a two-stage framework, TAC2, for timestamped audio captioning. Stage 1 addresses the data scarcity problem for fine-grained temporal grounding by programmatically fusing predictions from weak teachers (sound event detectors, diarizers, captioners) using a time-registered IoU matching algorithm. This creates a structured, speaker-attributed timeline without expensive manual annotation. The model is trained using an Instruction-Input-Output (IIO) completion format, which allows for composable functions where outputs of one task can condition subsequent tasks. Stage 2 extends this by eliciting latent paralinguistic capabilities (stress, emotion, etc.) from a frozen backbone using lightweight LoRA specialists trained on small datasets, then distilling these into a unified student model. The methodology is sound, leveraging the strengths of weak supervision and parameter-efficient fine-tuning to build a complex, structured representation.
The authors validate the model in two settings: audio understanding (cascading with frozen LLMs) and audio generation (paired supervision). In understanding, TAC2 achieves strong results on MMSU and MSU-Bench, outperforming baselines like TAC and Qwen3Omni-Captioner, particularly in perception tasks. The generation experiments demonstrate zero-shot control over event timing and speaker attributes. The evaluation is comprehensive, covering both the utility of the structured representation for reasoning and its application in generative tasks.
The paper provides detailed descriptions of the fusion algorithm, the IIO format, and the specialist training procedure. However, specific hyperparameters, dataset sizes for the weak supervision corpus, and the exact implementation details of the IoU matching thresholds are not fully specified in the provided text, which may hinder exact reproduction. The use of standard tools like LoRA and common benchmarks aids reproducibility.
The reliance on weak teachers means the quality of the fused timeline is bounded by the accuracy of the individual detectors and diarizers. Errors in the weak supervision stage can propagate to the final model. Additionally, the "latent capability elicitation" approach, while efficient, may not capture all nuances of paralinguistic features if they are not strongly encoded in the backbone. The paper is truncated, so the full extent of the generation results and ablations is not visible.
This work has significant potential for applications requiring precise audio understanding, such as automated video captioning, audio search, and accessible media. By decoupling perception from reasoning, it enables more robust and controllable audio pipelines. The composable nature of the IIO format makes it adaptable to various downstream tasks. TAC2 introduces a scalable framework for timestamped audio captioning by fusing weak teacher predictions and eliciting latent paralinguistic capabilities, resulting in a composable, temporally grounded representation that enhances both audio understanding and controllable generation.
Reinforcement learning has made Audio Large Language Models (Audio LLMs) better reasoners, but the reward that drives it still comes from outside the model: a final-answer check, a hand-crafted process score, or an external LLM judge that stays fixed while the policy improves. We show that an Audio LLM can supply this reward itself, and can do so recursively: each improved model becomes the critic that trains its successor, so the training signal improves along with the model. We introduce AudioRSI (Recursively Self-Improving Audio Reasoners), a reinforcement learning framework in which a single model plays two roles: as an actor it generates reasoning paths, and as a critic it compares them pairwise, turning its own judgments into a process-level reward. The critic is periodically replaced by the current policy, which closes the recursive loop. Because the critic hears the same recording as the actor, it can reject reasoning that cites acoustic evidence the audio does not contain, rather than reward reasoning that merely sounds plausible. The recursion is what sustains progress: under a frozen external judge, out-of-distribution accuracy plateaus during training, while a critic refreshed from the improving policy keeps it rising. Across four out-of-distribution benchmarks, all disjoint from the AVQA training distribution, AudioRSI outperforms strong baselines including Audio Flamingo 3 and Gemini 2.5 Pro on MMAU test-mini, ranks first on MMSU, achieves a clear lead on the challenging Multi-Audio split of MMAU-Pro, and leads the 7B open-source tier on MMAR among methods without external reward models. With no reward specifying how to reason, AudioRSI also develops structured reasoning behaviors (logical reasoning, perception-aware analysis, and reduced hallucination), corroborated by an external LLM-as-Judge evaluation and a keyword-based behavior analysis.
Primary: Amazon
All Institutions: University of Illinois Urbana-Champaign (UIUC), Amazon
AudioRSI introduces a recursively self-improving reinforcement learning framework for audio LLMs where the model serves as its own evolving critic. By periodically synchronizing the critic with the actor, the method ensures that the reward signal for reasoning quality improves alongside the policy, leading to superior out-of-distribution performance and emergent structured reasoning behaviors compared to static external judges or outcome-only rewards.
The paper proposes AudioRSI, a framework where an Audio LLM acts as both actor and critic within a GRPO (Group Relative Policy Optimization) loop. The core innovation is the "recursive" aspect: the critic is periodically synchronized with the actor's weights, allowing the reward signal to improve as the model improves. The methodology is sound, leveraging the unique advantage of audio modality where the critic can verify acoustic claims against the raw waveform, unlike text-only self-rewarding models. The use of pairwise comparison for process-level rewards is a solid adaptation of DPO-style thinking to RLVR (Reinforcement Learning with Verifiable Rewards).
The experiments are conducted on Qwen2.5-Omni-7B using AVQA for training and four OOD benchmarks (MMAU, MMSU, MMAU-Pro, MMAR) for evaluation. The results show consistent improvements over baseline RL methods (Ke-Omni-R) and strong performance against proprietary models like Gemini 2.5 Pro on specific splits. The ablation on synchronization period $T$ is crucial and well-executed, demonstrating the trade-off between stability and reward freshness. The claim that the model develops structured reasoning behaviors is supported by LLM-as-Judge and keyword analysis, though these are somewhat proxy metrics.
The paper provides detailed algorithmic descriptions and hyperparameters in the appendix. The reliance on specific base models (Qwen2.5-Omni) and datasets (AVQA) makes reproduction feasible for groups with similar compute resources. The code is not explicitly linked in the provided text, which is a minor gap for immediate reproducibility.
The approach is computationally expensive due to the dual role of the model (generating and critiquing). The "guarded handoff" mechanism, while practical, adds complexity. The evaluation relies heavily on LLM-as-Judge for reasoning quality, which may introduce bias if the judge model shares biases with the actor. The gains are modest in some domains (e.g., Music), suggesting the method may be more effective for domains with clearer acoustic evidence.
This work pushes the boundary of self-improving AI systems in multimodal domains. It demonstrates that models can provide meaningful process-level supervision for their own reasoning, reducing reliance on external, static reward models. This has implications for scaling audio reasoning capabilities without massive human annotation of reasoning steps. AudioRSI introduces a recursively self-improving reinforcement learning framework for audio LLMs where the model serves as its own evolving critic. By periodically synchronizing the critic with the actor, the method ensures that the reward signal for reasoning quality improves alongside the policy, leading to superior out-of-distribution performance and emergent structured reasoning behaviors compared to static external judges or outcome-only rewards.