This week’s shortlist
Music production typically involves multiple editing operations that each require specialized tools and expertise. Existing text-guided music editing systems support few operations or instruments, operate on short clips, or handle only one operation per instruction, unlike real…
Abstract
Music production typically involves multiple editing operations that each require specialized tools and expertise. Existing text-guided music editing systems support few operations or instruments, operate on short clips, or handle only one operation per instruction, unlike real workflows where several changes are requested at once. Here, we present ProCraft, a lightweight 0.5B-parameter autoregressive codec language model that edits 30s music excerpts end-to-end from free-form instructions, including compound requests. ProCraft supports seven operations across eight instruments, from tasks with a unique target (e.g., track removal) to creative ones (e.g., accompaniment generation). We align intermediate representations with pretrained music embeddings and pitch and instrument targets to preserve content outside the edit, and gated cross-attention in every decoder layer to keep generation tied to the instruction. To train on compound requests, we introduce On-the-fly Editing-pair Synthesis, which generates paired instructions and audio combining track-level operations with effect changes. ProCraft performs competitively against task-specific systems on single-task editing, and outperforms four free-text editors on compound instructions for track removal, extraction and denoising on real (MoisesDB) and synthetic (Slakh) recordings. In a user study (16 listeners), ProCraft was rated highest on removal and denoising and on-par with a 16-times larger model on extraction and instrument transfer. Ablations further show that representation alignment and cross-attention reduce leakage from unwanted tracks and raise editing success rates, especially on out-of-domain data.
AI-generated analysis. Check claims against the paper.
Summary
ProCraft introduces a lightweight autoregressive codec language model capable of performing diverse and compound music editing tasks from free-form text instructions. The paper makes a significant contribution by addressing the challenge of preserving unedited content while applying multiple simultaneous edits, using novel architectural components like a persistent edit head and intermediate layer alignment, along with a synthetic data generation pipeline that enables training on compound instructions.
Methodology
The paper proposes ProCraft, a 0.5B parameter autoregressive codec language model for music editing. The methodology is sound, leveraging the InspireMusic backbone and WavTokenizer. The introduction of a "persistent edit head" with gated cross-attention to maintain instruction relevance over long sequences is a logical architectural choice for autoregressive models. The "On-the-fly Editing-pair Synthesis" is a clever data generation strategy using a tool-calling LLM to create compound instruction-audio pairs, addressing a significant data scarcity issue in this niche. The alignment of intermediate layers with MERT and pitch/instrument targets is a strong technical contribution to preserve unedited content.
Results & evaluation
The evaluation is comprehensive, covering seven operations across eight instruments. The use of both synthetic (Slakh) and real (MoisesDB) data for out-of-domain testing is rigorous. The metrics are well-chosen, combining general audio metrics (FAD, CLAP) with task-specific ones (SI-SDR, edit rate, preservation metrics). The user study, while small (16 listeners), provides necessary subjective validation. The model outperforms baselines on compound instructions, which is the core claim.
Reproducibility
The paper promises to release code and weights. The training procedure, including the two-pass training for the edit head, is described in detail. The data synthesis pipeline is explained, though the specific prompts and tool-calling logic for the LLM are likely in the appendix. The use of standard components (WavTokenizer, MERT) aids reproducibility.
Limitations
The model is limited to 30-second excerpts, which is short for full music production workflows. The user study is small. The model is trained primarily on synthetic data (Slakh) and tested on real data, which may limit generalization to complex, real-world mixes with non-standard instruments or production styles. The 0.5B parameter size, while efficient, may limit the complexity of creative edits compared to larger models.
Broader impact
This work contributes to the field of controllable audio generation by demonstrating that a single lightweight model can handle diverse and compound editing tasks. It bridges the gap between task-specific models and general-purpose text-guided editors. The data synthesis technique is applicable to other audio editing tasks. ProCraft introduces a lightweight autoregressive codec language model capable of performing diverse and compound music editing tasks from free-form text instructions. The paper makes a significant contribution by addressing the challenge of preserving unedited content while applying multiple simultaneous edits, using novel architectural components like a persistent edit head and intermediate layer alignment, along with a synthetic data generation pipeline that enables training on compound instructions.
Streaming video-to-audio (V2A) generation is essential for immersive interactive applications. However, real-time constraints leave existing streaming V2A methods far behind offline models: (1) visual frontends shrink their receptive fields to cut look-ahead latency, severely…
Abstract
Streaming video-to-audio (V2A) generation is essential for immersive interactive applications. However, real-time constraints leave existing streaming V2A methods far behind offline models: (1) visual frontends shrink their receptive fields to cut look-ahead latency, severely degrading audio-visual alignment; (2) causal autoregressive diffusion backends denoise each audio patch in isolation, impairing audio quality. To this end, we present LiveFoley, a streaming causal autoregressive diffusion architecture that addresses both bottlenecks with imperceptible latency. For the visual frontend, we introduce a Trailing Sliding Window that converts non-causal segment-level encoders into streaming modules without retraining. By simply realigning the window, it preserves precise audio-visual alignment, while a configurable hop size reduces the frontend look-ahead well below the human perceptual threshold. Building on this tunable hop size, we propose the Unified Varhop Framework to accommodate scenarios with diverse latency and computation budgets. For the audio backend, we explore widening the denoising window of the local diffusion transformer (LocDiT) head in two temporal directions. History LocDiT adds past context to improve quality at no latency cost and serves as our default. Alternatively, Future-Warmup LocDiT breaks the autoregressive deadlock with an approximate future condition, further improving generation quality at the cost of an acceptable backend look-ahead. Extensive experiments demonstrate that LiveFoley matches offline models in temporal alignment and leads streaming systems in generation quality. Project page: https://dalision.github.io/livefoley-page/
AI-generated analysis. Check claims against the paper.
Summary
LiveFoley introduces a streaming causal autoregressive diffusion architecture that achieves imperceptible latency and high-quality audio-visual alignment by decoupling frontend look-ahead from receptive field size via a Trailing Sliding Window and enhancing backend denoising with historical and future-warmup contexts. The technical contribution is significant for the field of real-time multimodal generation, offering a practical solution to the latency-quality trade-off that has hindered the adoption of high-fidelity V2A models in interactive environments.
Methodology
The paper proposes LiveFoley, a streaming video-to-audio generation system that addresses the latency-quality trade-off in real-time applications. The core contribution is the "Trailing Sliding Window" mechanism, which allows non-causal segment-level encoders (like Synchformer) to operate in a streaming context without retraining by anchoring the window tail to the current frame and using first-frame padding. This decouples look-ahead latency from the receptive field size. Additionally, the "Unified Varhop Framework" introduces a Synchronizer module to handle variable hop sizes, allowing the system to adapt to different latency budgets dynamically. For the audio backend, the authors modify the Local Diffusion Transformer (LocDiT) by introducing "History LocDiT" (using past context) and "Future-Warmup LocDiT" (using approximate future conditions) to improve generation quality compared to isolated patch denoising. The methodology is sound and logically addresses the identified bottlenecks in existing streaming V2A systems.
Results & evaluation
The paper includes a user study to quantify human perceptual thresholds for audio-visual delay, which is a valuable contribution to the field. Experiments demonstrate that LiveFoley matches offline models in temporal alignment while maintaining low latency. It claims to lead streaming systems in generation quality. The use of a single A800 GPU for real-time performance is a strong practical result. However, the text provided is truncated, so the full extent of quantitative comparisons (e.g., FAD, CLAP scores) against specific baselines like V-AURA or SoundReactor is not fully visible, though the abstract and intro suggest comprehensive testing.
Reproducibility
The paper provides architectural details and a project page. The use of standard components (MiniCPM, Synchformer, flow-matching) aids reproducibility. The specific implementation details of the Synchronizer and the training protocol for the variable hop sizes are described, but code availability is not explicitly confirmed in the provided text (only a project page URL is given).
Limitations
The system relies on a causal LM backbone which may still struggle with very long-range dependencies compared to global bidirectional models. The "Future-Warmup" mechanism introduces a small amount of additional latency (0.08s) which, while below perceptual thresholds, adds complexity. The evaluation is primarily focused on latency and alignment; subjective audio quality metrics (MOS) are likely included in the full paper but are not detailed in the truncated text.
Broader impact
This work has significant potential for immersive interactive applications such as AR/VR, video games, and live streaming, where real-time audio feedback is crucial. By providing a plug-and-play sound engine that works with arbitrary visual pipelines, it lowers the barrier for integrating high-quality synchronized audio into interactive media. LiveFoley introduces a streaming causal autoregressive diffusion architecture that achieves imperceptible latency and high-quality audio-visual alignment by decoupling frontend look-ahead from receptive field size via a Trailing Sliding Window and enhancing backend denoising with historical and future-warmup contexts. The technical contribution is significant for the field of real-time multimodal generation, offering a practical solution to the latency-quality trade-off that has hindered the adoption of high-fidelity V2A models in interactive environments.
Simultaneous Speech Translation (SimulST) requires balancing translation quality and latency. While Large Language Models (LLMs) excel in offline translation, current LLM-based SimulST systems predominantly rely on prefix-incremental generation. This paradigm suffers from…
Abstract
Simultaneous Speech Translation (SimulST) requires balancing translation quality and latency. While Large Language Models (LLMs) excel in offline translation, current LLM-based SimulST systems predominantly rely on prefix-incremental generation. This paradigm suffers from redundant Key-Value cache recomputation and necessitates external Voice Activity Detection to process unbounded streams. To overcome these bottlenecks, we propose an end-to-end interleaved read-write framework built upon an omni-modal LLM. By interleaving speech chunks and text tokens, our architecture achieves full KV cache reuse, natively supporting segmentation-free, infinite-length inference. Furthermore, to learn optimal streaming policies without rigid offline alignments, we introduce a novel Actor-Critic self-enhancement pipeline. An RNN-T Actor dynamically samples read-write trajectories, while an LLM-based Critic evaluates these paths via intrinsic semantic and latency rewards, distilling the optimal policy back to the Actor. Extensive experiments on sentence-level (FLEURS, CoVoST2) and document-level (MuST-C, ACL 60/60) benchmarks demonstrate that our model establishes new state-of-the-art quality-latency Pareto frontiers, substantially outperforming existing cascaded and static-alignment baselines.
AI-generated analysis. Check claims against the paper.
Summary
The paper proposes CRESiST, an end-to-end interleaved read-write framework for simultaneous speech translation that leverages an omni-modal LLM and an Actor-Critic self-enhancement pipeline to optimize streaming policies without rigid offline alignments. This work represents a strong technical contribution to the field of LLM-integrated audio processing, addressing key bottlenecks in latency and cache efficiency for infinite-length inference. While it does not reach the level of a paradigm-shifting landmark like the Transformer or BERT, it offers a significant architectural improvement over existing prefix-incremental methods, establishing new state-of-the-art Pareto frontiers for quality and latency. The integration of reinforcement learning (Actor-Critic) for policy distillation in a speech translation context is a novel and meaningful advance, likely to be cited by researchers working on streaming speech models and real-time translation systems.
A speech deepfake detector estimates whether a recording is synthetic, but a forensic analyst also asks which system produced it, whether recordings share a source, and what evidence supports each conclusion. We present VeriSpoof, an audio language model that unifies five tasks…
Abstract
A speech deepfake detector estimates whether a recording is synthetic, but a forensic analyst also asks which system produced it, whether recordings share a source, and what evidence supports each conclusion. We present VeriSpoof, an audio language model that unifies five tasks with shared weights: deepfake detection, generator attribution (generator, acoustic model and vocoder), recording-condition analysis, synthesized-span localisation, and source verification. Each decision is supported by grounded rationales whose statements can be independently verified. Answers are scored by sentence likelihood, allowing detection and source verification to be thresholded without separate heads. Rationale statements can be checked against waveform measurements, span labels, or generator metadata. Counterfactual examples train decisions to follow the stated evidence. Across the 14 Speech-DF-Arena test sets, VeriSpoof-Plus achieves 2.14% pooled EER, ranking third of 28 systems and first among open-weight systems. VeriSpoof-Base, trained only on official training splits of eight corpora, achieves 4.43% EER, less than half that of the next open-weight detector. With the same weights, VeriSpoof-Plus verifies whether recordings from generators unseen in training share a source at 6.2% EER. Controlled studies show that the audio representation governs detection, the LLM governs recording comparison, and source supervision is necessary for neural-codec fakes. Statement-checkable rationales catch partial fakes missed by one-step answers (99.6% vs. 78.4% on PartialEdit) and determine decisions under edits unseen in training (99.5%). Training updates only 29.7M of 2.85B parameters, fitting a single 24 GB GPU with gradient checkpointing. We release both models, inference code, a demo, and VeriSpoof-Bench, the evaluation suite covering these questions.
AI-generated analysis. Check claims against the paper.
Summary
VeriSpoof unifies multiple speech forensics tasks into a single audio language model with verifiable, grounded rationales. The paper makes a substantial contribution by demonstrating that LLMs can perform complex forensic reasoning (attribution, source verification) with high accuracy when equipped with checkable intermediate steps, outperforming specialized detectors on open-weight benchmarks and providing a new standard for explainable AI in audio security.
Methodology
The paper proposes VeriSpoof, an audio language model (ALM) that unifies five distinct forensic tasks (deepfake detection, generator attribution, recording-condition analysis, synthesized-span localisation, and source verification) into a single architecture. The core methodological innovation is the "grounded rationale" framework, where the model generates explanations composed of fixed-field statements (e.g., SNR, bandwidth, span labels) that can be independently verified against waveform measurements or metadata. This addresses the "faithfulness" problem in LLM explanations by training with counterfactual examples that force the decision to depend on the stated evidence rather than just the audio input. The model uses a Qwen3-ASR encoder and MiniCPM5-2B LLM, employing parameter-efficient fine-tuning (LoRA/DoRA) on only 29.7M of 2.85B parameters. The scoring mechanism reads the likelihood of specific answer sentences, eliminating the need for separate classification heads for different tasks.
Results & evaluation
The experimental evaluation is extensive and rigorous. On the Speech-DF-Arena benchmark (14 test sets), VeriSpoof-Plus achieves a pooled EER of 2.14%, ranking third overall and first among open-weight systems. VeriSpoof-Base, trained on fewer data, achieves 4.43% EER, significantly outperforming other open-weight detectors. The paper includes controlled ablation studies to determine the contribution of the encoder vs. the LLM, showing that the audio representation governs detection while the LLM governs source comparison. It also demonstrates that source supervision is critical for detecting neural-codec fakes. The rationale mechanism is shown to catch partial fakes (99.6% vs 78.4% for one-step answers) and maintain decision faithfulness under unseen edit types (99.5%).
Reproducibility
High. The authors release both model weights (Base and Plus), inference code, a demo, and a comprehensive evaluation suite (VeriSpoof-Bench) with 177,852 items. The training recipe, data composition, and hyperparameters are detailed. The claim that training fits on a single 24GB GPU with gradient checkpointing is supported by memory analysis, making the work accessible to a broader research community.
Limitations
The model relies on specific base models (Qwen3-ASR, MiniCPM5-2B) which may not be the most optimal for all audio tasks. The "grounded rationale" approach requires careful construction of checkers for each statement type, which may not generalize automatically to new forensic questions without manual checker design. The performance on unseen languages (SpoofRadar) shows a calibration cost where pooled metrics hide per-set threshold issues. The paper is a preprint (arXiv) and has not yet undergone peer review at a major conference.
Broader impact
This work has significant implications for digital forensics and media integrity. By providing explainable, verifiable, and multi-functional deepfake detection, it moves the field beyond simple binary classification toward actionable forensic analysis. The ability to attribute generators and verify sources with grounded evidence could aid law enforcement and platform moderation. The release of open-weight models and benchmarks lowers the barrier to entry for developing forensic audio tools. VeriSpoof unifies multiple speech forensics tasks into a single audio language model with verifiable, grounded rationales. The paper makes a substantial contribution by demonstrating that LLMs can perform complex forensic reasoning (attribution, source verification) with high accuracy when equipped with checkable intermediate steps, outperforming specialized detectors on open-weight benchmarks and providing a new standard for explainable AI in audio security.
Speaker-recognition benchmarks test how accurately a model recovers the speaker label of a recording, treating voice identity as a property of the person who produced it. However, these benchmarks do not evaluate how humans perceive voice identity. Whether a listener hears two…
Abstract
Speaker-recognition benchmarks test how accurately a model recovers the speaker label of a recording, treating voice identity as a property of the person who produced it. However, these benchmarks do not evaluate how humans perceive voice identity. Whether a listener hears two voices as the same person has consequences for intellectual property rights, cybersecurity, and everyday life. We introduce the Voice Identity Perception benchmark (VIPBench), which contains 124,876 same-or-different judgments from 1,290 listeners on 9,800 recording pairs from 100 public figures. The pairs include real recordings, voice clones generated by a text-to-speech system, and morphs between two speakers' voices. We also define four tasks that evaluate the alignment between speaker-representation models and human identity perception. The first two tasks measure how well a model's similarity score predicts the proportion of listeners who respond "same" and the majority response. The other two measure whether a model and listeners find the same speaker pairs similar, and whether a model's relationship to listener judgments holds equally for real and synthetic recordings. We evaluate eleven publicly available speech models. Listener judgments diverge from speaker labels most often for voice clones: in 41% of responses, listeners judged a clone and a real recording of the same speaker to be different people. No model reaches the noise ceiling; the best model explains 61% of the explainable variance. Models trained with speaker labels align with listeners better than self-supervised models do, yet a model with a 12% error rate on speaker labels aligns with listeners as well as models with error rates below 2%. VIPBench enables the evaluation of speaker models against human perception on real, cloned, and morphed speech, and our results motivate training speaker models directly on listener judgments.
AI-generated analysis. Check claims against the paper.
Summary
The paper introduces VIPBench, a comprehensive benchmark for evaluating the alignment between speech representation models and human perception of voice identity. It makes a significant contribution to the field by providing a large-scale dataset of human judgments on real, cloned, and morphed voices, along with a rigorous evaluation protocol that reveals the limitations of current speaker verification models in capturing human perceptual nuances.
Methodology
The paper introduces VIPBench, a large-scale benchmark (124,876 judgments) designed to evaluate how well speech representation models align with human perception of voice identity, rather than just metadata-based speaker labels. The methodology is rigorous, involving a stratified sampling of 100 public figures across sociophonetic groups, genders, and ages. The stimuli include real recordings, state-of-the-art voice clones (Cartesia Sonic 3), and continuous morphs. The evaluation protocol defines four distinct tasks: continuous prediction of perceived sameness, Platt-calibrated verification, speaker-level representational similarity, and real-to-synthetic calibration transfer. The use of cross-validation over speakers and the inclusion of a noise ceiling based on inter-rater reliability demonstrates a sophisticated statistical approach to benchmarking.
Results & evaluation
The experiments evaluate 11 publicly available models, including supervised (x-vector, ECAPA-TDNN, Resemblyzer) and self-supervised (wav2vec 2.0, HuBERT, WavLM, XLS-R) models, as well as Whisper. The results reveal that while supervised models generally outperform self-supervised ones in aligning with human perception, none reach the noise ceiling. A key finding is the significant divergence between metadata labels and human perception for voice clones, where listeners often perceive clones as different speakers. The paper provides detailed ablations on layer selection for SSL models and robustness checks using unfamiliar trials and different TTS systems (CosyVoice 2). The inclusion of demographic fairness analysis is a strong addition, showing consistent performance across sociophonetic groups for most models.
Reproducibility
The paper is highly reproducible. The dataset, code, and embeddings are released on Hugging Face. The authors provide detailed information on stimulus generation, including the specific TTS settings and random seeds used for morphing. The analysis notebook is provided, allowing users to reproduce the results without needing to re-extract embeddings. The clear documentation of the IRB protocol and data collection process further enhances trust and reproducibility.
Limitations
The benchmark is limited to American English speakers and public figures, which may not generalize to other languages or private individuals. The reliance on a single commercial TTS system (Cartesia) for the main set of clones and morphs could introduce system-specific biases, although a small extension with CosyVoice 2 mitigates this partially. The lack of test-retest reliability for individual listeners is a minor limitation, though the high inter-rater reliability compensates for this.
Broader impact
This work has significant implications for the development of voice biometric systems, cybersecurity, and intellectual property protection. By highlighting the gap between metadata-based speaker verification and human perception, the paper motivates the training of speaker models directly on human perceptual judgments. This could lead to more robust and human-aligned voice authentication systems that are better equipped to handle voice cloning attacks. The benchmark also provides a valuable resource for researchers studying human voice perception and the alignment of machine representations with human cognition. The paper introduces VIPBench, a comprehensive benchmark for evaluating the alignment between speech representation models and human perception of voice identity. It makes a significant contribution to the field by providing a large-scale dataset of human judgments on real, cloned, and morphed voices, along with a rigorous evaluation protocol that reveals the limitations of current speaker verification models in capturing human perceptual nuances.
Dialectal Arabic speech recognition remains constrained by the scarcity of dialect-resolved transcribed speech despite the abundance of web audio. We introduce GigaSpeech 3, a corpus of 39,000 hours and 10 million weakly transcribed speech segments covering Algerian, Egyptian…
Abstract
Dialectal Arabic speech recognition remains constrained by the scarcity of dialect-resolved transcribed speech despite the abundance of web audio. We introduce GigaSpeech 3, a corpus of 39,000 hours and 10 million weakly transcribed speech segments covering Algerian, Egyptian, Iraqi, Saudi, and Syrian Arabic. Starting from 60 hours of manual transcription per dialect, we train labeling models using speech-recognition-biased self-supervised learning and apply them to dialect-matched web audio. Data selection and four rounds of noisy-student refinement for Saudi and Syrian Arabic yield the final corpus. We validate its utility through dialect-specific and multi-dialect speech recognition training and cross-domain evaluation. On GigaSpeechBench, the final dialect-specific models outperform their labeling models across all five dialects and achieve a macro-average word error rate of 23.86%, a 16.6% relative reduction from the strongest third-party system covering all five dialects, while outperforming the best third-party result on three dialects. A multi-dialect model using self-predicted dialect identities achieves a macro-average word error rate of 25.51% while also supporting Modern Standard Arabic. Across eight conditions from three external corpora, a GigaSpeech 3-trained Zipformer achieves the lowest word error rate among six evaluated models in every condition, while adapting the Qwen3 speech recognizer on GigaSpeech 3 improves performance across all eight conditions. We release the corpus and recipes as a foundation for dialect-aware Arabic speech processing.
AI-generated analysis. Check claims against the paper.
Summary
The paper presents GigaSpeech 3, a large-scale weakly labeled corpus for five Arabic dialects, constructed via ASR-biased self-supervised learning and noisy-student refinement, demonstrating that such weak supervision enables compact models to outperform strong third-party systems and large pretrained models in dialect-specific and multi-dialect recognition tasks.
Methodology
The paper proposes a data-centric approach to low-resource dialectal Arabic ASR. The core methodology involves constructing a large-scale weakly labeled corpus (GigaSpeech 3) using a "VietASR" pipeline. This pipeline leverages ASR-biased self-supervised learning (SSL) to train labeling models from small manual seeds (60 hours/dialect), which are then applied to dialect-matched web audio. The process includes data selection and iterative Noisy Student Training (NST) refinement for specific dialects (Saudi, Syrian). The novelty lies not in a new model architecture, but in the rigorous application of existing SSL and weak-labeling techniques to a specific, under-served linguistic domain (Arabic dialects) at a significant scale (39k hours). The validation strategy is sound: training independent downstream models (Zipformer) from scratch on the weak labels to prove the utility of the data beyond the labeling model's parameters.
Results & evaluation
The experimental evaluation is comprehensive and rigorous. The authors validate the corpus by showing that models trained on the weak labels outperform the labeling models themselves, confirming the data's utility. They compare against strong third-party baselines (including commercial APIs and large open-source models like Qwen3 and Whisper) on GigaSpeechBench, achieving state-of-the-art results for compact models. Cross-domain evaluation on MASC, Casablanca, and SADA further demonstrates generalizability. The inclusion of a multi-dialect model with automatic language identification (LID) and the analysis of routing errors adds depth to the evaluation. The results are consistent and well-supported by ablation studies (e.g., Labeled-only vs. SSL+FT vs. Weak-labels).
Reproducibility
High. The authors release the code, dataset, and demo. The pipeline is described in detail, including the specific stages of the VietASR recipe and the NST refinement process. The use of standard architectures (Zipformer) and public benchmarks (GigaSpeechBench) enhances reproducibility. The separation of training and test channels at the source level is a strong methodological choice that is clearly documented.
Limitations
The primary limitation is the reliance on web audio, which may contain noise, code-switching, or non-dialectal content that is difficult to fully filter. The country-level dialect labels are coarse and may not capture intra-country variation. The multi-dialect model's performance is affected by LID errors, particularly for Egyptian and Syrian dialects, as shown in the confusion matrix. The paper focuses on five specific dialects, leaving other Arabic varieties unaddressed.
Broader impact
This work has significant impact on the field of low-resource speech recognition. By providing a large, dialect-resolved corpus and a validated pipeline for weak labeling, it enables the development of more accurate and inclusive ASR systems for Arabic speakers. The methodology can be adapted to other low-resource languages or dialects. The release of the corpus and tools will likely accelerate research in dialectal Arabic processing and serve as a benchmark for future work. The paper presents GigaSpeech 3, a large-scale weakly labeled corpus for five Arabic dialects, constructed via ASR-biased self-supervised learning and noisy-student refinement, demonstrating that such weak supervision enables compact models to outperform strong third-party systems and large pretrained models in dialect-specific and multi-dialect recognition tasks.
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…
Abstract
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.
AI-generated analysis. Check claims against the paper.
Summary
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.
Methodology
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.
Results & evaluation
The evaluation is comprehensive, testing a wide range of open-weight and proprietary models. The key finding—that open-weight models struggle with audio perception while frontier models excel at symbolic reasoning with CoT—is 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.
Reproducibility
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.
Limitations
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.
Broader impact
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…
Abstract
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/
AI-generated analysis. Check claims against the paper.
Summary
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.
Methodology
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.
Results & evaluation
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.
Reproducibility
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.
Limitations
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.
Broader impact
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.
Video dubbing aims to synthesize lip movements that match target speech while preserving the speaker identity, head pose, and background of a source video. Existing editors are predominantly trained via self-reconstruction under original paired audio, even though the source…
Abstract
Video dubbing aims to synthesize lip movements that match target speech while preserving the speaker identity, head pose, and background of a source video. Existing editors are predominantly trained via self-reconstruction under original paired audio, even though the source mouth becomes a conflicting target when driving speech is replaced. We introduce replacement-aware cross-audio supervision: paired-audio samples use full-frame reconstruction, whereas unpaired replacement-audio samples mask out source-mouth reconstruction losses and supervise the mouth solely through audio-visual synchronization while preserving unmasked facial regions and background. Because offline dubbing provides the complete source video and target speech beforehand, we present VidDub, a full-context diffusion editor distilled into two denoising steps (2 NFE). Our replacement-aware teacher first learns accurate lip synchronization without regressing toward conflicting source mouth movements. Matched trajectory regression then transfers the teacher's CFG-guided trajectories to a two-step student, and distribution matching distillation recovers visual quality after step compression. We will release the curated training dataset, VidDub-Bench, fixed paired- and cross-audio evaluation manifests, preprocessing outputs, and unified evaluation code. Across three training seeds, masking out conflicting source-mouth reconstruction improves cross-audio Sync-C by 0.833; the final two-step model combines strong lip synchronization with superior visual quality, identity preservation, and visual speech recognition across four benchmarks.
AI-generated analysis. Check claims against the paper.
Summary
The paper introduces a replacement-aware distillation framework for two-step video dubbing, effectively resolving the conflict between source mouth reconstruction and target audio synchronization through masked supervision. The technical contribution is significant, combining a novel training objective with efficient distillation to achieve state-of-the-art performance in lip synchronization and visual quality, supported by a comprehensive evaluation and the release of a new benchmark.
Methodology
The paper proposes a two-stage approach for video dubbing: a replacement-aware teacher training strategy and a two-step distillation pipeline. The core methodological contribution is the "replacement-aware cross-audio supervision," which addresses the conflict between source mouth reconstruction and target audio synchronization by masking out source-mouth reconstruction losses during training on unpaired audio. This is a logical and effective solution to a known problem in video editing models. The distillation process uses matched trajectory regression and distribution matching to compress a 48-step teacher into a 2-step student, which is a standard but well-executed application of recent diffusion distillation techniques. The architecture relies on Wan-1.3B with VACE conditioning, which is a solid baseline.
Results & evaluation
The experimental evaluation is comprehensive, covering multiple benchmarks (HDTF, TalkVid-Bench, VidDub-Bench) and metrics (Sync-C, Sync-D, FVD, SSIM, PSNR, VSR). The paper demonstrates state-of-the-art performance in cross-audio synchronization while maintaining high visual quality. The ablation studies clearly show the benefit of the masking strategy. The comparison with both feed-forward and multi-step diffusion methods is fair and informative. The release of a curated dataset and benchmark (VidDub-Bench) adds significant value to the community.
Reproducibility
The authors commit to releasing the training dataset, benchmark, preprocessing outputs, and unified evaluation code. The GitHub repository is provided. The detailed description of the training pipeline, loss functions, and distillation steps suggests high reproducibility, assuming the released code is complete and well-documented.
Limitations
The model is designed for offline dubbing, requiring the full source video and target speech beforehand, which limits its applicability to real-time or streaming scenarios. The reliance on a specific base model (Wan-1.3B) may limit generalizability to other architectures. The evaluation, while extensive, is primarily quantitative; subjective human evaluation (MOS) would further strengthen the claims about visual quality and naturalness.
Broader impact
This work has significant potential for applications in video localization, accessibility, and content creation. By enabling high-quality, fast (2-step) video dubbing, it can reduce the cost and time associated with traditional dubbing processes. The open-source release of the dataset and code will likely accelerate research in audio-visual synchronization and video editing. The paper introduces a replacement-aware distillation framework for two-step video dubbing, effectively resolving the conflict between source mouth reconstruction and target audio synchronization through masked supervision. The technical contribution is significant, combining a novel training objective with efficient distillation to achieve state-of-the-art performance in lip synchronization and visual quality, supported by a comprehensive evaluation and the release of a new benchmark.
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…
Abstract
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.
AI-generated analysis. Check claims against the paper.
Summary
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.
Methodology
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.
Results & evaluation
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.
Reproducibility
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.
Limitations
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.
Broader impact
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…
Abstract
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/.
AI-generated analysis. Check claims against the paper.
Summary
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.
Methodology
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.
Results & evaluation
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.
Reproducibility
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.
Limitations
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.
Broader impact
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…
Abstract
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
AI-generated analysis. Check claims against the paper.
Summary
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
Methodology
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.
Results & evaluation
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