Overview

Speech foundation models learn from large speech or audio-text collections, often using self-supervised methods such as masked prediction, then support tasks including automatic speech recognition, speaker analysis, emotion recognition, and text-to-speech. Common approaches include fine-tuning models such as wav2vec 2.0, HuBERT, WavLM, and Whisper; parameter-efficient adapters; and generative models built around discrete audio tokens or continuous speech latents. Researchers evaluate them on suites such as SUPERB, multilingual recognition tests, and benchmarks for robustness and speech generation. Models now span many tasks and languages, but performance can falter under noise, codec changes, domain shifts, and limited training data, leaving an open tension between broad generality, reliable adaptation, and efficient deployment.

Latest papers

71 papers5 papers in the last four weeks, down 64% on the four weeks before. 0.0% of all new papers.

Nov 10, 2026

Walrus Technical Report

We present Walrus, a unified 8B speech-and-audio foundation model handling speech recognition (including code-switching and multi-talker speech), speech-to-text and speech-to-speech translation, text-to-speech, emotion recognition, audio captioning, and general audio understanding. Our main contribution is not a new architecture but a data-centric training recipe; we use a standard decoder-only backbone initialized from Gemma 4, so that gains can be traced to the data. For speech generation, Walrus predicts low-bitrate discrete speech tokens, and a lightweight token-to-waveform decoder converts them into high-fidelity audio. The recipe is a three-stage data curriculum that moves from broad multilingual data to multi-talker, code-switching, and domain-specific data. At each stage, task-aware and dataset-aware weighting balances tasks with very different token budgets and protects core skills from forgetting. Trained on over 600M samples covering more than 20 tasks and 100 languages, Walrus leads on most multilingual, multi-talker, and code-switching ASR, speech translation, and audio understanding benchmarks, outperforming both open speech language models of comparable size and dedicated task-specific models. On most benchmarks, it even matches or surpasses the 4×4\times larger Qwen3-Omni's performance. Results suggest that data engineering is a primary driver of quality in multi-task speech and audio modeling.
Nov 10, 2026

Nepali-English Code-Mixed Speech Recognition: Data Preparation and Adaptation of a Speech Foundation Model

Nepali speakers regularly mix English words into conversation, but multilingual speech recognisers often struggle with language switches and the writing script used for English words. We study how to adapt SraVaani, a 449M-parameter hybrid TDT-CTC model, using a 172.8-hour corpus covering six Nepali speech domains. Its pretrained tokenizer does not support digits and several common symbols. Rather than replace it or discard these recordings, we generate Nepali and English readings for numbers and use zero-shot ASR output to select a plausible form. Full fine-tuning reduces test word error rate (WER) from 48.6% to 17.1%; a 12-epoch run reaches 16.3%, compared with 16.4% for fine-tuned Whisper large-v3-turbo. Freezing the encoder increases WER to 25.5%, while excluding numeric clips raises WER on numeric test speech from 14.3% to 21.3%. We also evaluate on FLEURS using both its original references and mixed-script alternatives. The comparison shows that transcription conventions explain part, but not all, of the out-of-domain performance loss.
Nov 5, 2026

Route-DMD: Distribution Matching Distillation with Task-Routed Mechanism for Speech Foundation Models

Speech foundation models built on flow matching handle generation, editing, and processing in one system, but every request costs 20-50 iterative steps, and classifier-free guidance (CFG) doubles each of them. Distribution matching distillation (DMD) trains a student that approaches the teacher's quality in a few steps, yet existing speech distillation work applies it to generation alone. We show that a distribution-matching objective does not serve every task: in separation, the noised states of erroneous student outputs deviate far from the teacher's training distribution, where its prediction is unreliable. Route-DMD therefore makes the objective task-dependent, routing the affected tasks to a regression loss and keeping decoupled distribution matching elsewhere. The student is initialized by consistency distillation, and teacher rollouts use adaptive projected guidance (APG) instead of CFG. Without CFG, the 4-step student matches the 32-step teacher on generation, editing, and processing tasks at a 4.5×\times end-to-end speedup.
Nov 3, 2026

TTSonic: A Unified Framework for Composable and Controllable Speech Generation

Modern text-to-speech (TTS) systems can synthesize natural and expressive speech, yet a practical speech foundation model must go beyond naturalness to offer stable quality, precise controllability, and a broad set of composable capabilities within a unified framework. We present TTSonic, a unified speech generation model that supports zero-shot TTS, voice cloning, prosody cloning, instruction following, voice conversion, and cloning from noisy reference speech. Rather than relying on separate task-specific models, TTSonic accommodates diverse conditioning signals, including text, natural language instructions, reference speech, and text-described timbre, within the same model framework. Specifically, we develop a semantic--acoustic disentangled speech tokenizer through ASR-supervised semantic learning and reconstruction-guided generative refinement. Serving as the representational foundation of our unified framework, the tokenizer enables the Text-to-Semantic autoregressive model to focus on semantic and prosodic planning, while the acoustic flow-matching model specializes in high-quality speech rendering under diverse acoustic and timbre conditions. Building on this framework, we incorporate CoT-guided instruction-following planning and Text2Timbre voice design to support controllable speech generation. Experiments across zero-shot synthesis, multilingual and cross-lingual TTS, instruction following, voice design, voice conversion, and noisy-reference cloning demonstrate the competitive performance and broad applicability of TTSonic.
Oct 28, 2026

Pronunciation-Supervised Text-to-Speech for Irish (Gaeilge)

Speech technology for Irish must be built from very little transcribed speech, yet the language's pronunciation is thoroughly documented in dictionaries and dialect descriptions. We ask whether that documentation can supervise the adaptation of a large pretrained text-to-speech model to a language it was never trained on. We adapt Qwen3-TTS to Irish using 12.8 hours of crowdsourced Common Voice speech, correcting two defects in the vendor fine-tuning script that training loss does not reveal (Appendix A), and reach character error rate (CER) 0.225 on held-out cross-corpus text, against 0.397 for an exhaustively tuned from-scratch system and 0.437 for zero-shot cloning. We then compile the Wiktionary pronunciation record into a lexicon and a phoneme critic, and use that critic as the verifier in one round of rejection fine-tuning: CER 0.159 cross-corpus and 0.067 in-domain, with the voice still anonymous across all 202 training speakers. The same critic explains why the voice cannot yet be steered to a minority dialect: 229 of 12,311 training clips lean Ulster against 8,451 Munster. Pronunciation documentation, converted into a verifier, is a practical lever for low-resource speech synthesis, and a way to measure what the available data cannot support.
Oct 18, 2026

How Far Can the Gemma 4 Speech Encoder Go Without Changing Its Weights? Saudi and Gulf Arabic Speech Recognition with a Frozen and an Adapter-Augmented Encoder

We study the 305M-parameter audio encoder shared by Gemma 4 and EmbeddingGemma 2 (bit-identical to that of Gemma 4 E2B except for the final audio-embedding projection) for Saudi and Gulf Arabic speech recognition, without changing any of its pretrained weights; the adapters of our adapted system are additive and kept separate. A frozen system trains only a 22.6M-parameter CTC reader over a learned mix of four intermediate encoder layers and the encoder output; an adapted system adds 3.1M low-rank adapter parameters, a 41.5M-parameter hybrid CTC/attention reader and joint decoding with a sub-word language model. Against Cohere Transcribe Arabic (2.07B parameters), re-run on identical audio, the frozen system trails in word error rate (WER) but is not significantly different in character error rate (CER) on five of six human-labelled conditions, while running 3.4 times faster on CPU with greedy decoding. The adapted system is not significantly different in WER on a public Saudi customer-service call (19.20% against 19.27%) and on clean Emirati speech, is significantly better on telephone-coded Emirati speech, has a significantly lower CER on all six conditions of these sets, and beats a 0.6B recogniser fine-tuned on SADA in WER on all six; it still trails Cohere in WER on long-form Saudi speech. On a blind, human-corrected public call-centre sample it ranks 4 of 15 systems, behind three commercial APIs (significantly only one) and ahead of the other self-hostable systems by point estimate. An optimised joint decoder runs it at 139 times real time on one GPU. We also find, on our data, that references made by correcting machine drafts favour the drafting system. Code and public-set results are released.
Oct 14, 2026

Audio System One Model: Typed Probabilistic Perception of Emotion in Speech

Speech applications ask many questions about the same voice: which emotion a listener perceives, how aroused the speaker sounds, whether they sound frustrated. Classifiers answer fast, over one fixed label set; audio language models (ALMs) take any question in text, but answer by generation, one pass per question, with poorly calibrated probabilities. We present ASOM (Audio System One Model), a System One model for speech: choice, score and yes/no questions written at request time are answered in one pass, without generation, each with a probability distribution. Frozen speech and text encoders feed a 6.2M-parameter reader with no per-class parameters, trained on listener vote distributions. On MSP-Podcast, ASOM matches a classifier on the same features and data, leads the best of 6 zero-shot ALMs by 0.072 and an ALM fine-tuned on the same data by 0.026, and answers 20 questions in 13.1 ms, 17-41 times faster than the ALMs. It also answers reworded, reduced and new label sets, scales and gender, and on 21 unseen corpora in 13 languages it leads the best zero-shot ALM by 0.138 where the words carry no emotion; where they do, the ALMs mostly read.
Oct 13, 2026

Code-Switching Detection via Language-Selective Dimensions in Speech Foundation Models

Code-switching (CS), the alternation between languages within an utterance, is common in multilingual speech but underrepresented in pretraining corpora. Increasing CS coverage requires finding relevant utterances in large speech collections so that they can be included in training data. One way to find such utterances is to transcribe the speech and detect language alternation in the text, but this depends on obtaining accurate multilingual transcriptions. We propose a lightweight method that detects CS directly from speech. It uses frame-level language information in pretrained speech encoders to determine whether both languages are present within an utterance. Specifically, we identify language-selective hidden-state dimensions using monolingual recordings and aggregate their frame-level activations into a CS prediction. With a small CS set used to tune a decision rule, we show that the Whisper-small encoder achieves a mean F1 of 0.7980.798 across 14 language pairs in CS-FLEURS, compared with 0.7760.776 for a CNN and 0.7510.751 for a HuBERT linear baseline trained for monolingual language identification. Our findings suggest that pretrained speech encoders already encode the information needed to identify CS utterances, and they provide a lightweight tool for building more representative multilingual corpora.
Oct 13, 2026

Fish Audio Drama 3 Technical Report

We introduce Fish Audio Drama 3, an open-source speech generation model that turns text into directed performances, built on a 35B-parameter mixture-of-experts backbone that activates 3B parameters per token. It clones voices from multiple references, follows natural-language direction, revises a take when the next turn gives a note, voices multi-character scenes, edits existing speech by infilling, and synthesizes long-form and streaming audio. Multi-turn revision, scene dubbing, editing by infilling, and streaming text input are new in Drama 3, all in one model. Drama 3 pairs the hybrid linear-attention backbone with a new codebook multi-token-prediction audio decoder that predicts the acoustic codebooks of each frame in dedicated layers that attend over the entire sequence. The model is trained with a three-stage pre-training recipe, instruction fine-tuning, and large-batch reinforcement learning whose only reward is a single learned model of human preference. Its inference engine decodes a single stream at 271 frames per second on one H200, against 177 for Fish Audio S2, although Drama 3 has more than seven times as many parameters. The model weights are available on GitHub (https://github.com/fishaudio/fish-speech) and Hugging Face (https://huggingface.co/fishaudio/drama-3), and inference is available through SGLang-Omni (https://github.com/sgl-project/sglang-omni). We highly encourage readers to visit https://fish.audio to direct their own performances.
Oct 12, 2026

Nufo: A Benchmark for Ewe Automatic Speech Recognition

Automatic speech recognition (ASR) remains challenging for low-resource languages, where the scarcity of transcribed speech limits the development of accurate systems. This challenge is particularly acute for African languages such as Ewe, which has millions of speakers but remains underrepresented in publicly available ASR resources and benchmarks. In this work, we consolidate, normalize, and partition existing Ewe speech resources into a standardized 174-hour Ewe ASR benchmark, the largest to date to our knowledge. We fine-tune and evaluate pretrained speech models, including Whisper and AfriHuBERT, to establish strong baselines on our benchmark. Whisper-medium achieves a word error rate (WER) of 37.98%, while AfriHuBERT achieves 39.70% on the test set. We further provide a qualitative analysis of common recognition errors. We release the benchmark under the name Nufo, derived from the Ewe word for "talk", accompanied by models and evaluation resources, to facilitate future research on Ewe and other underrepresented African languages.
Oct 7, 2026

Backdooring Acoustic Foundation Models for Physically Realizable Triggers

Acoustic foundation models (AFMs) have democratized acoustic applications, enabling powerful models for tasks ranging from speech recognition to speaker verification with minimal resources. However, the security of applications based on AFMs remains largely underexplored. Our work addresses this gap by proposing the Foundation Acoustic model Backdoor (FAB) attack, demonstrating that state-of-the-art AFMs are susceptible to backdooring under practical settings. Despite making minimal assumptions about adversary capabilities (e.g., no access to pre-training data), we show that FAB preserves benign performance while inducing backdoors that survive fine-tuning and cause significant degradation across diverse downstream tasks when activated. Notably, FAB utilizes task-agnostic, physically realizable, inconspicuous, and sync-free triggers (e.g., a background siren). We evaluate FAB using two leading AFMs, nine downstream tasks, and four different triggers. We further demonstrate its effectiveness against established defenses and across both digital and physical domains. While extensive end-to-end fine-tuning can mitigate FAB, such a defense is resource-intensive and task-specific. Our work highlights critical risks to AFMs and calls for advanced defenses.
Oct 6, 2026

BanglaBox: A Phonetically-Balanced Corpus and Data-Efficient Foundation-Model Adaptation for Bangla Text-to-Speech with Zero-Shot Voice Cloning

We present a recipe for adapting English-pretrained autoregressive TTS foundation models to underrepresented languages, demonstrated on Bangladeshi Bangla. Existing Bangla TTS corpora are small and single-speaker, and to our knowledge no open zero-shot voice-cloning system is available for the Bangladeshi register. We contribute a phonetically- and gender-balanced two-tier Bangladeshi Bangla corpus balanced via a tiered Jensen-Shannon divergence objective over conjunct clusters (juktakkhor), together with three fine-tuning changes: a merge-consistent tokenizer extension, Bangla text normalization, and a prompt-masked dual-loss objective. These changes preserve zero-shot cloning across the language switch. Our BanglaEval protocol applies Wilcoxon signed-rank tests with Bonferroni correction over native-speaker ratings. BanglaBox attains near-natural Naturalness, outperforms commercial and open-source baselines on Naturalness, Speaker Similarity, and Clarity, and reaches speaker similarity comparable to prior few-shot results while using approximately 7x less Bangla fine-tuning audio. We further validate the recipe beyond our own test split using a seven-category stress set of difficult real-world text, the public BnTTS evaluation benchmarks, and naturally occurring Bangla that no language model wrote. All artifacts, including the corpus, weights, code, and complete evaluation materials, are released publicly and unconditionally.
Oct 5, 2026

Neural Representations, Natural Connections: What Transfers From Human Speech Foundation Models to Animal Vocalizations?

Speech-pretrained models have shown promise in animal bioacoustics, but the factors governing their cross-species transfer remain poorly understood. We evaluate 15 frozen encoders spanning monolingual and multilingual speech, speaker verification, animal bioacoustics, and general audio on caller identification across four species and call-type classification across three. Differences between the strongest speech- and animal-pretrained representations range from −0.027-0.027 to +0.119+0.119 UAR; speech is significantly better in three of seven settings (p<0.01p<0.01) and not significantly different in the remaining four. Controlled comparisons show no systematic advantage from increased language coverage or animal-domain pretraining, while frozen speaker-verification embeddings transfer poorly. The results also show that transfer is strongly layer-dependent: raw-waveform models peak early, whereas patch-spectrogram models peak deeper.
Oct 4, 2026

Transferable Adversarial Robustness for Speech Foundation Models via Hierarchical Stabilization

Frozen speech foundation models (SFMs) make downstream adaptation efficient: the backbone can stay fixed while a task learns layer fusion and a lightweight classifier. Full adversarial fine-tuning is a standard route to robustness, but generating adversarial examples and updating the backbone for every task sacrifices that efficiency. We ask whether robustness can instead be learned before future tasks are known. For a frozen backbone and linear classifier, robustness can be understood through the interaction between representation stability and decision-boundary margin. This leads directly to our design: we stabilize representations across the hidden layers, rather than only the final layer, while preserving clean representations; after clean adaptation selects the layer mixture, we keep it fixed and enlarge only the classifier margin, without downstream adversarial examples. We evaluate Wav2Vec2, HuBERT, and WavLM Large on four tasks under adaptive 30 dB attacks. Across 12 backbone-task pairs, hierarchical robustification improves robust accuracy by 46.4 pp, while margin refinement adds 4.0 pp for 1.1 pp of clean accuracy. Code and configurations are available at https://github.com/arefmousavi/hierarchical-robust-sfm.
Sep 24, 2026

Adaptive Fisher-Whitened Cross-Covariance for Low-Resource Speech Recognition

Adapting multilingual speech foundation models to low-resource languages remains difficult, especially for languages that are poorly represented during pre-training. While parameter-efficient fine-tuning (PEFT) reduces the cost of adapting large models, conventional approaches such as LoRA rely on generic low-rank parameterizations and do not explicitly use downstream task information to define the adaptation subspace. To investigate whether task-informed PEFT can better support low-resource ASR, we apply Fisher-Whitened Cross-Covariance Analysis (FCCA) to Whisper and Qwen3-ASR, and introduce two complementary extensions: Asymmetric-Coupled FCCA (AC-FCCA), which exploits structured cross-layer sharing, and Adaptive-Rank FCCA (AR-FCCA), which reallocates adaptation capacity across projection matrices under a fixed parameter budget. Under controlled multilingual experiments, we evaluate these approaches on languages that are poorly represented or unsupported during pre-training alongside well-represented languages. Standard FCCA is competitive with, and usually outperforms, trainable-parameter-budget-matched LoRA. AR-FCCA provides the most consistent improvement over standard FCCA across both model architectures, with statistically significant gains in several evaluation settings, while retaining the same number of trainable parameters. These results show that task-informed subspace construction can be effective for low-resource speech adaptation, and that adaptive rank allocation provides a robust way to improve parameter efficiency without increasing model capacity.
Sep 23, 2026

Temporal Taxation Compounds Under Post-Training Compression of Whisper Models

Automatic speech recognition models are audited for demographic fairness at full precision, yet the models that ship to production have been quantized, pruned, and distilled. We ask whether post-training weight compression, which alters model weights rather than the audio signal or its feature representation, redistributes error burden across demographic groups. Across the Whisper family on Fair-Speech, Common Voice 25, and AfriSpeech-200, 50% Wanda pruning of Whisper-large-v3 sharply widens the Black/AA-vs-Asian temporal-taxation differential on Fair-Speech: the absolute word-error-rate gap between the worst- and best-served groups more than doubles; at an assumed cost of five seconds of correction effort per transcription error this is a rise from 30 to 64 seconds of correction time per minute of speech. This +111% relative increase is invariant to the assumed per-error cost, survives an audio-quality control, and is only partly mitigated by beam-search decoding, which still leaves an +86% increase. At edge model size, INT4 HQQ quantization compounds catastrophic transcript loops on West African accents by factors of five to seven. Distillation, by contrast, narrows demographic gaps in 21 of 27 evaluated settings (teacher-student pair, precision, and dataset), with the exceptions concentrated on a single model pair. We cast the temporal-taxation construct of Choi and Choi (2025) as a quantitative metric, and show that single-snapshot fairness audits on full-precision models do not capture the deployment-time burden that compression places on already-marginalized speakers.
Sep 23, 2026

Six Layers Less: Encoder Pruning for Whisper with Label-Free Recovery

Pruning large pre-trained transformer-based ASR models such as OpenAI's Whisper has seen great adoption, as pruning the decoder led to significant end-to-end transcription speedups. For instance, the {\tt whisper-large-v3-turbo} variant reduced the decoder from 32 to 4 layers, while Distill-Whisper similarly reduced the decoder to only 2 layers. Although some attention has been put towards reducing the size of the encoder, no approach has seen wide adoption. This could be due to the need for custom inference implementations to take advantage of the compressed model. We present an approach that ranks encoder layers by the leave-one-layer-out change in Word Error Rate (WER). The six layers that cause the least change are removed, corresponding to 18.5%18.5\% of the encoder stack. The pruned model requires no custom inference code as it is simply a more shallow encoder with fewer layers. We further distill using unlabeled monolingual speech data to recover performance degradation caused by the zero-shot layer pruning. Mean WER across four languages increases to 20.1%20.1\% after distillation, compared to 21.9%21.9\% zero-shot, going from a baseline of 18.2%18.2\%. We release all of our code (https://github.com/rasgaard/whisper-encoder-layer-prune) and the pruned model (https://huggingface.co/rasgaard/whisper-large-v3-turbo-encoder-pruned).
Sep 21, 2026

Qwen-Audio-3.1-Realtime: Towards Reliable Agentic Voice Interaction

Real-time voice assistants must reason over evolving requests, execute actions, and follow conversational rules. Qwen-Audio-3.1-Realtime brings these requirements together through Think, Act, and Speak and Coordinate. Think combines Core-Cocktail supervised fine-tuning with Multimodality and Multi-Teacher On-Policy Distillation (M2^{2}-OPD) to transfer language capabilities and develop native audio skills. Act uses self-evolving executable environments and multi-granularity rollouts for Group Relative Policy Optimization (GRPO), teaching the model to use tools, interpret feedback, and complete tasks. Speak and Coordinate aligns whether, when, and how the assistant speaks or acts. We evaluate audio reasoning, multilingual understanding, tool use, conversational behavior, full-duplex interaction, and safety. Compared with Qwen-Audio-3.0-Realtime, 3.1 raises overall task success from 78.4% to 82.0% on our half-duplex speech-to-text adaptation of ττ-Voice. On speech-to-speech Full-Duplex-Bench v1.5, the response rate to background speech falls from 73.0% to 13.0%. We also present a separate Voice Harness prototype, using Qwen-Audio-3.0-Realtime as its foreground, that extends spoken interaction to persistent tasks through foreground--background coordination and memory.
Sep 21, 2026

AURA: Uncertainty-Routed Activation Editing for Acoustic Grounding in Speech Foundation Models

Attention encoder-decoder (AED) Speech Foundation Models achieve strong ASR performance but can generate acoustically unsupported text when inputs contain no speech, weak acoustic evidence, or unreliable transcription. We propose AURA: Activation-editing with Uncertainty-Routed Adaptation, an ultra-efficient representation-editing method that freezes the pretrained model and applies sparse scale-and-shift edits to decoder cross-attention heads. AURA dynamically routes edits using cross-attention uncertainty features that capture over-concentration, diffuse attention, and abrupt frame shifts. We evaluate AURA on four datasets spanning non-speech hallucination and speech grounding stressors, including imperfect-label child speech, imperfect-label adult speech, and disfluent speech. On non-speech audio, AURA reduces hallucination rate from 89.18% to 1.94% without prior hallucination-head identification. On imperfect-label corpora, AURA approaches LoRA WER while using roughly 500x fewer trainable parameters. Sensitivity analysis and qualitative cross-attention examples are consistent with AURA's uncertainty-routed editing behavior, supporting dynamic activation editing as a practical path for grounding AED speech models.
Sep 16, 2026

Cross-Lingual Parkinson's Disease Severity Assessment Using Pre-trained Speech Embeddings: A Multi-Class Evaluation

Parkinson's disease (PD) often manifests through speech impairments, facilitating accessible, non-invasive, and cost-effective severity assessment for early diagnosis and progression tracking. Despite advances in speech foundation models (SFMs), their cross-lingual generalization for PD severity multi-class classification remains underexplored due to limited labeled data, a lack of explainable methods and variability across languages and datasets. In this work, we evaluate pre-trained embeddings from four state-of-the-art open-source SFMs across three datasets in zero-shot and k-shot cross-lingual settings for multi-class PD severity assessment. Our results show that pre-trained speech embeddings enable meaningful cross-lingual transfer, although performance is sensitive to dataset properties, preprocessing, and adaptation strategy. Misclassifications under these conditions related to inter-speaker variability and atypical speech patterns highlight the need for more robust feature extraction and modeling for PD severity assessment while emphasizing the importance of explainability for reliable clinical insights.
Sep 14, 2026

Listening for Airway Stenosis: A Foundation Model-Based Method for Rapid and Accessible Detection

Airway stenosis can cause severe respiratory complications, yet its detection often relies on specialized examinations and medical imaging. This study explores the potential of acoustic AI for rapid and accessible airway stenosis detection using readily acquired patient voice recordings. We systematically investigate whether acoustic foundation models (AFMs) can extract acoustic representations associated with airway stenosis-related speech patterns. Experiments are conducted on a cohort of 748 participants from the Bridge2AI-Voice dataset, 134 with airway stenosis and 614 without. The best-performing model achieves an AUROC of 0.952 and an accuracy of 0.924 (means over five-fold cross-validation), highlighting the potential of AFMs to transfer beyond general-purpose speech applications to clinical diagnostic tasks. Further analysis reveals that the model primarily relies on connected-speech recordings rather than isolated acoustic tasks, such as sustained phonation and breathing. Overall, these results suggest that voice-based acoustic AI could complement existing diagnostic workflows by enabling rapid, low-burden, and widely accessible screening for airway stenosis.
Sep 14, 2026

CAL-MOS: Bridging Layers with Adapters for Robust MOS Prediction Across Speech Foundation Models

Speech Quality Assessment (SQA) is essential for modern speech technologies, and recent non-intrusive SQA predictors increasingly rely on Speech Foundation Models (SFMs). However, because SFMs expose representations from many layers, it remains unclear which depths are most informative for MOS prediction and how multi-layer information should be combined reliably across backbones and datasets. We benchmark ten SFMs on four MOS datasets under three regimes: full fine-tuning, last-layer probing with a frozen encoder, and naive cross-layer weighted aggregation. We find that the best layer is strongly backbone- and dataset-dependent, and that naive weighted fusion can be unstable across settings. We further evaluate a layer-calibrated aggregation variant that applies per-layer adapters before pooling, which improves the robustness of multi-layer fusion and narrows the gap to full fine-tuning while keeping the backbone frozen.
Sep 9, 2026

Do speech foundation models really learn words?

Self-supervised speech foundation models are now used in a wide array of downstream applications, including traditional speech recognition and as the basis for tokens in speech-aware language models. Attempts to understand their usefulness have largely focused on probing their representations' ability to discriminate phonemes and words. However, discriminative ability for words need not imply specialized representation of words per se. Good discrimination of words may be explained by good encoding of word form (phonemes) rather than form-independent word representations encoding identity or syntactic/semantic properties. By partialling out phoneme information using residualization, we show that, in later layers, HuBERT and wav2vec 2.0 do in general learn representations which encode words with reasonable fidelity independently of local phonetic content. We show that this simple approach to disentanglement can enhance higher-order linguistic information in word discovery tasks.
Sep 9, 2026

Sparse Weight and Edge Circuit Discovery in Transformer-based Acoustic Models

Transformer-based foundation models are powerful but opaque, motivating Mechanistic Interpretation methods to uncover the black-box by identifying small computation subgraphs responsible for a task. DiscoGP is a joint weight-and-edge circuit discovery framework originally developed for text decoders. We extend DiscoGP to speech encoders and present, to our knowledge, the first circuit discovery study for modern speech foundation models. Across HuBERT and Wav2Vec 2.0 on several speech classification tasks, we find that the discovered circuits are extremely compact, yet often match or even exceed the performance of the full pretrained encoder with the same downstream head. Through ablations, we show that these circuits reflect pretrained computation rather than random structure or task-head artifacts. We also introduce a memory-efficient DiscoGP variant that reduces the GPU memory cost of edge-circuit discovery at runtime from quartic to cubic. Overall, our results broaden Mechanistic Interpretation beyond text decoders and show that circuit-level analysis can reveal both explanatory structure and unexpected functional behavior in speech encoders.
Sep 9, 2026

BuzzASR: A Swarm of 100+ Monolingual Speech Recognition Models

We introduce BuzzASR, a collection of language-specialized fine-tuned Whisper models adapted for automatic speech recognition (ASR) in 102 languages. Large end-to-end Transformer-based ASR models such as Whisper have revolutionized ASR, but most prominent models are highly multilingual. As a result, these models often perform poorly on languages less well-represented in their training set. While it has long been known that effective language adaptation can be achieved through simple fine-tuning on monolingual data, this strategy has only been applied to a small number of languages. We massively scale up this simple approach to 102 languages covered in the FLEURS dataset, while also implementing a more complex language adaptation strategy that integrates monolingual tokenizer replacement and data augmentation using text-only fine-tuning. BuzzASR models outperform Whisper-large-v3 on 77 out of 102 languages, reducing character error rates (CER) by a factor of over 2.8 on average. Our models achieve state-of-the-art CER among open-source systems on 27 of 102 languages on the combined FLEURS and Common Voice test set. Our tokenizer replacement strategy yields an average 3.3x improvement in compression rate (characters per token) over Whisper's multilingual BPE, with gains of up to 21.7x. We release all models, code, and detailed results: https://lemn-lab.github.io/buzz-asr
Sep 8, 2026

AuK Technical Report: An Open-Source Foundational Model for Speech Generation and Editing

We introduce AuK, an open-source foundational model that unifies speech generation and editing through a common interface of natural-language instructions and audio context. To support this broad capability set, we construct approximately 3.03 billion instruction--audio instances and 1.95 million hours of effective supervision across five task families: speech generation, content editing, enhancement and separation, paralinguistic editing, and acoustic editing. AuK combines a multimodal large language model for semantic conditioning, an VAE jointly trained on speech, general audio, and music for acoustic conditioning, and a hybrid rectified-flow Transformer that performs dual-stream MMDiT blocks followed by unified single-stream DiT blocks for generation. Training begins with generation-only warm-up and proceeds to joint generation--editing pre-training. We then apply complementary post-training strategies: human-feedback preference optimization for open-ended editing and reward-based reinforcement learning for speech generation. To reduce inference cost, we further distill the model with consistency initialization and task-routed Decoupled DMD. The resulting AuK-Flash performs 4-step inference without classifier-free guidance and achieves a 4.5 wall-clock speedup over the full model under matched conditions. Experiments demonstrate leading performance on zero-shot and instruction-controlled speech generation and general instruction-guided editing, while remaining competitive on signal-level restoration tasks. We release both the source code and model weights to support reproducibility and further research.
Sep 7, 2026

Clean Accuracy Does Not Guarantee Provenance Robustness: A Prospective Codec-Stress Evaluation of Audio Attribution

Audio provenance attribution - which system produced a synthetic utterance - is reported at near-ceiling accuracy on clean benchmarks, yet audio reaching an analyst has usually been transcoded. We report a prospectively registered measurement of closed-set attribution after single-stage codec transport, with the analysis region fixed from fidelity metadata before any attribution model was trained. On two corpora, in-support losses reach 53.5 [43.5, 63.6] and 70.3 [63.0, 77.5] Macro-F1 points for WavLM-Base+, and 61.0 [56.8, 65.1] and 49.8 [41.6, 57.9] for W2V2-BERT 2.0, under simultaneous component-level bands. Degradation is strongly condition- and representation-dependent: within one in-support grid WavLM losses run from -0.4 to +53.5 points, and the two encoders differ beyond a prespecified +/-5-point margin at six of twelve conditions. A clean-qualified ECAPA-TDNN and a Proxy-Anchor head degrade comparably, so the effect is not confined to one representation family or a weak linear head. The registered matched-fidelity comparison was not estimable on this grid, and waveform and perceptual measures order the conditions differently: MP3 at 8 kbit/s ranks mid-grid on SI-SDR but last on PESQ-WB while causing the largest loss. For the tested tasks, corpora, representations and codec grid, a clean accuracy figure does not by itself characterise deployment robustness.
Sep 4, 2026

KanAdapter: A Kolmogorov-Arnold Network-based Plug-and-Play Module for Efficient Fine-tuning of Foundation Speech Models

Fully fine-tuning self-supervised learning (SSL) speech models for downstream tasks is computationally prohibitive, and existing parameter-efficient fine-tuning approaches predominantly rely on MLP-based adapters whose fixed activation functions limit their representational expressiveness under tight parameter budgets. We propose KanAdapter, a lightweight adapter framework that replaces conventional MLP bottlenecks with Group-Rational Kolmogorov-Arnold Network (GR-KAN) modules for more expressive and parameter-efficient adaptation. Following a parallel bottleneck design, KanAdapter inserts trainable GR-KAN branches alongside frozen Transformer encoder blocks and leverages weight transfer from pre-trained MLP layers for stable initialization. Across speaker verification, speech emotion recognition, and deepfake detection, KanAdapter achieves up to 97.5% reduction in trainable parameters relative to full fine-tuning while remaining highly competitive, and consistently outperforms AdaptFormer under comparable parameter budgets. In continual learning, it yields up to 83.6% error reduction over full fine-tuning and MLP-based adapters, which we attribute to the localized nature of GR-KAN's rational activations that mitigates catastrophic forgetting. To our knowledge, this is the first work to explore KAN-based modules for parameter-efficient fine-tuning of speech foundation models.
Sep 3, 2026

Reducing Hallucinated Transcripts in Whisper via Hallucination Space Projection

Whisper is a widely used foundation model for automatic speech recognition (ASR), but its generative decoder can produce fluent hallucinated transcripts for inputs containing little or no speech. We propose a training-free, inference-time method to reduce these hallucinations using low-rank projection of decoder activations. A compact hallucination-associated subspace is estimated from non-speech calibration data, and decoder hidden states are projected away from this subspace during inference. We evaluate two variants: always-on, which applies projection to all inputs, and gated, which applies it only when Whisper predicts that an input is likely non-speech. Across non-speech benchmarks, always-on projection reduces average hallucination rate (HR) from 31.31% to 2.44%, a 92.21% relative reduction, while gated projection reduces HR to 3.74%, an 88.05% relative reduction, with lower false rejection of genuine speech. On LibriSpeech, gated projection increases absolute word error rate (WER) by 0.33-4.39 percentage points and yields false-rejection rates (FRR) of 0.41--9.97% across model and split settings. These results show that low-rank activation projection can substantially suppress Whisper hallucinations without retraining, while providing a controllable trade-off between hallucination suppression and speech recognition performance.
Aug 26, 2026

Fine-Tuning Whisper for Automatic Speech Recognition in Baniwa: A Preliminary Study

Automatic Speech Recognition (ASR) technologies have achieved remarkable performance in recent years through the use of large multilingual foundation models. However, most advances remain concentrated on high-resource languages, while indigenous languages continue to suffer from a lack of speech resources and language technologies. This work presents a preliminary study on the adaptation of Whisper for Automatic Speech Recognition in Baniwa, an indigenous Arawakan language spoken in Brazil, Colombia, and Venezuela. The experiments were conducted using a corpus of 1,373 manually transcribed recordings obtained from a linguistic documentation project. The corpus contains approximately 0.54 hours of speech and consists primarily of isolated words and short elicited utterances. The Whisper Small model was fine-tuned using supervised learning and evaluated using Word Error Rate (WER) and Character Error Rate (CER). The best model achieved a WER of 37.5% and a CER of 7.45%, demonstrating that multilingual foundation models can be successfully adapted to extremely low-resource indigenous languages. The results establish an initial baseline for Baniwa Automatic Speech Recognition and provide a foundation for future research involving larger datasets, language-specific adaptation strategies, and post-processing techniques.