Papers for

speech technology engineers

Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.

Image models reused to create clearer and faster speech synthesis

ImageTTS: Repurposing Pretrained Image Diffusion Transformers for Speech Synthesis

Abstract: Pretrained image-generation models have been successfully repurposed for various downstream visual tasks. Whether their learned representations and generative priors can transfer beyond vision to other modalities remains largely unresolved, particularly for speech synthesis, which requires linguistic fidelity and preservation of speaker identity from a short reference. We present ImageTTS, which repurposes FLUX.2 for speech synthesis by encoding pseudo-RGB log-mel spectrograms with its frozen image VAE and adapting only its pretrained Diffusion Transformer (DiT) using conditional latent flow matching. This formulation disentangles two forms of vision-to-speech transfer: representation transfer through the image VAE and generative-prior transfer through DiT initialization. The frozen image VAE largely preserves intelligibility and speaker identity with 0.99 STOI and 0.95 speaker similarity, while image-pretrained DiT initialization accelerates convergence and improves linguistic accuracy, reducing WER from 4.2\% to 2.7\% at 100k matched updates relative to random initialization. The same framework supports both zero-shot and instruction-guided TTS and, when scaled to 100K hours of speech data, achieves 1.12\% WER on Seed-TTS test-en and a Style-ACC of 0.66 on CapSpeech, outperforming speech-native baselines trained on comparable data across both tasks. These results suggest that pretrained image models can transfer beyond architectural reuse, supplying useful latent representations and generative priors for intelligible and controllable speech synthesis.

Mon 26 OctSoundMultimedia
The gist
Making computers produce speech that sounds like a particular person is tricky and needs a lot of training. This paper shows it is possible to use parts of artificial intelligence models trained on pictures to help make speech. The authors keep some image-processing parts frozen and only retrain certain components to make spoken words that are clearer and sound like the original speaker. They show this approach works better and faster than starting from scratch and can even follow voice instructions or use different speaking styles.
Open → 2610.34807v1

Using pseudo-labels as rewards improves speech recognition accuracy

Pseudo-Labels as Rewards, Not Targets: Group-Relative Policy Optimization for Semi-Supervised ASR

Abstract: Adapting ASR to a domain with only untranscribed audio typically relies on pseudo-labeling, a strong teacher transcribes the audio and a student is fine-tuned on it. Because supervised fine-tuning (SFT) treats every pseudo-label token as a target, the student inherits the teacher's errors. We show that using the same pseudo-labels as a reward reference for Group Relative Policy Optimization (GRPO) is more robust to label noise. In the banking domain, GRPO yields 29.3% to 56.2% relative WER improvements over SFT on Whisper-generated labels, and the advantage holds across four Whisper sizes, four teacher families, three domains and four languages. While SFT reproduces almost 100% of teacher errors verbatim, GRPO passes through only 27%. Rollout statistics explain why, erroneous pseudo-label words rarely appear in on-policy rollouts, so group-relative advantages stay aligned with the clean transcript, whereas SFT forces them into the policy's support and collapses the reward variance GRPO relies on.

Fri 23 OctComputation and Language
The gist
Training speech recognition systems usually relies on labels telling the model what was said, but getting these is expensive. People create 'pseudo-labels' by using an existing system to guess the words, then teach a new system to copy those guesses. The problem is if the guesses contain errors, the new system simply learns those mistakes. The authors show a new way to use these guesses not as exact answers but as rewards that guide learning, which helps the new system avoid copying errors and improves accuracy.
Open → 2610.32572v1

Search methods improve efficiency in speech recognition systems

A Systematic Comparison of Search Space Complexity and Efficiency for Time-Synchronous End-to-End ASR Models

Abstract: Search is a central component of Automatic Speech Recognition (ASR) systems and has been deeply investigated for classical ASR modeling based on Hidden Markov Models (HMMs). However, in work on current sequence-to-sequence ASR architectures it is often simplified to a small fixed-size beam search without external Language Model (LM). In this work, we present a comparative analysis of search space complexity and efficiency for Connectionist Temporal Classification (CTC) and neural transducer models under different pruning strategies and LM configurations. We show that these end-to-end (E2E) models work with a small search space most of the time but exhibit large temporary peaks in hypothesis count during periods of uncertainty. Score-based pruning dynamically adapts the search to model uncertainty, leading to a significantly lower Real-Time Factor (RTF). Our results highlight the limitations of search with low fixed-size beam pruning and provide practical guidelines for efficient search in E2E ASR.

Thu 22 OctComputation and Language
The gist
Automatic speech recognition systems turn spoken words into text, but how they search through possible interpretations can slow them down. The authors studied different search methods used in modern end-to-end speech recognition models, showing that fixed small searches miss important possibilities when the system is uncertain. Using a score-based method that adapts the search size dynamically helps the system work faster without losing accuracy. Their work suggests better ways to balance search complexity and speed in speech recognition.
Open → 2610.31259v1

Dynamic fusion improves few shot audio classification accuracy

DF-PALM: Dynamic Fusion Prompt Learning for Audio--Language Models

Abstract: Few-shot audio classification requires audio--language models to distinguish variable acoustic events from limited labeled data. PALM's main formulation adapts encoded class representations without changing pretrained encoders, but its class representations remain input-independent and classification uses audio features without class-conditioned refinement. We introduce DF-PALM, a dynamic feature-space prompting framework that combines audio-conditioned class representations with asymmetric cross-attention for audio refinement. Class representations are conditioned after text encoding, held fixed across fusion layers for each input, and reused for final scoring. Both pretrained encoders remain frozen, avoiding audio-conditioned text-encoder passes. With PENGI as the frozen backbone, the main 16-shot evaluation covered 11 benchmarks and six baselines. DF-PALM achieved 80.69% mean accuracy and 80.32% macro-averaged F1, exceeding PALM by 4.11 and 4.30 percentage points, respectively. Paired benchmark-level tests supported both gains after Holm correction. Across four shot settings, DF-PALM achieved the highest mean accuracy and macro-averaged F1 among the four evaluated prompt-learning methods. The gains extended to LAION-CLAP and OpenFLAM, with significant improvements over PALM in both metrics. At 50 candidate classes, operator-counted inference floating-point operations increased by 0.39% relative to PALM. These results support instance-conditioned feature-space adaptation as an effective approach to few-shot audio classification. Our code is publicly available at GitHub https://github.com/sunyulin0421/DF-PALM.

Sun 18 OctSoundMachine Learning
The gist
Classifying sounds with only a few labeled examples is challenging. The authors introduce DF-PALM, which better adjusts sound and text features together for each audio instance without retraining existing encoders. This method improves accuracy and reliability in recognizing different sound events compared to previous approaches. Their approach works well across many sound datasets using limited examples.
Open → 2610.24921v1

Speech encoder learns multiple voice traits at once with multi-teacher method

Learning Speech Attribute Representations via Multi-Teacher Distillation

Abstract: Recent post-training approaches have shown that self-supervised speech encoders can be adapted to learn a single utterance-level speech attribute through knowledge distillation. In this work, we investigate whether a single speech encoder can simultaneously preserve multiple complementary utterance-level speech representations. We propose a multi-attribute multi-teacher distillation framework that learns semantic, speaker, language, and emotion representations from four frozen teacher models using a shared SSL backbone and attribute-specific branches. Experiments on multilingual semantic retrieval, speaker verification, language identification, and speech emotion recognition show that the proposed model achieves performance close to dedicated single-task students while requiring only a single shared encoder. Furthermore, the learned layer aggregation strategies reveal that different speech attributes naturally emerge at different depths of the SSL encoder, providing new insights into the organization of speech representations.

Thu 15 OctComputation and Language
The gist
Speech systems usually learn one thing at a time, like understanding words or recognizing the speaker. This paper explores whether one system can learn many voice features at once, such as what is said, who is speaking, the language, and the emotion. The authors use a method where a main system learns from four experts specializing in each trait simultaneously. The result is a single speech encoder that performs nearly as well as systems trained only for one task, but much more efficiently. They also found that different voice traits are best recognized at different levels inside the system, offering clues about how speech information is organized.
Open → 2610.21941v1

Lossless method speeds up speech recognition by nearly two times

Lossless Inference Acceleration for ASR via EAGLE-Style Speculative Decoding

Abstract: Autoregressive Transformer and Conformer models have achieved state-of-the-art performance in automatic speech recognition (ASR). However, their sequential token generation mechanism imposes a significant bottleneck during decoding, resulting in substantial inference latency. While speculative decoding has emerged as a promising lossless acceleration technique, existing ASR-oriented approaches either suffer from low draft acceptance rates due to suboptimal draft quality or rely on specially designed external draft models, which limit their generality and practicality. To overcome these limitations, we propose EAGLE-ASR, a novel framework that accelerates ASR inference via EAGLE-style speculative decoding while strictly preserving lossless output quality. Specifically, we introduce Speech-Draft, a lightweight autoregressive module that serves as a universal draft model by jointly leveraging encoder acoustic features and hierarchical decoder representations from the target ASR model. Furthermore, we adopt a multi-step generation strategy with self-distilled data to train the Speech-Draft, enabling it to better approximate the target model's output distribution and maintain causal consistency during multi-token generation. Extensive experiments on representative Whisper and Conformer AED models demonstrate that EAGLE-ASR achieves significant lossless speedups of up to $1.9\times$ and $2.4\times$, respectively, consistently outperforming existing speculative decoding baselines for ASR. Our proposed framework further generalizes effectively to multilingual ASR and speech translation tasks, establishing EAGLE-ASR as a practical and universally applicable solution for accelerating modern ASR systems without compromising recognition accuracy.

Fri 9 OctSound
The gist
Automatic speech recognition models are very good but slow because they generate words one at a time. The authors introduce a new lightweight draft system called Speech-Draft that guesses multiple words ahead, helping the main model work faster without making mistakes. This new method, called EAGLE-ASR, speeds up recognizing speech by up to 2.4 times while keeping perfect accuracy. It works well on different popular speech models and even supports multiple languages and speech translation.
Open → 2610.13506v1

Comparing what you analyze matters more than how you analyze it

A Better Metric Will Not Save You: What You Compare Beats How You Compare It

Abstract: Representational similarity measures (RSA, CKA, Procrustes) are chosen for the transformations they treat as irrelevant, and all of them take the set of conditions as given. That set is fixed by two decisions, which inputs are admitted and how they are grouped, both taken before the matrix a measure operates on exists: no invariance property reaches them. We measure what they are worth where the arrangement the conditions ought to have is known independently of the systems compared, the acoustic organisation of vowels, across fifteen languages and twenty-seven representations. Changing which categories are compared, at a fixed number of them, moves the gap between two representations by more than twice the effect being measured, while changing how many moves it not at all. In this setting rankings of models are identified under these choices, and rankings across languages are not.

Fri 9 OctComputation and Language
The gist
When comparing different systems, it's common to use metrics to see how similar they are. This paper finds that the choice of which items or groups you compare has a bigger effect on the comparison result than the specific metric you use. The authors tested this on vowel sounds across many languages and found that changing comparison groups changes results more than changing the number of groups or the metric itself. This means how you decide what to compare is more important than how you measure it.
Open → 2610.13476v1

Yorùbá contour tones clarified with new writing marks

Marking Contour Tones in Yorùbá

Abstract: Yorùbá is a tonal language in which contour tones pose persistent orthographic challenges. These are especially notable for personal names and lexical items whose conventional spellings avoid vowel lengthening that would otherwise provide a host syllable for the second tone. A particular concern is a class of names in which the conventional spelling does not just omit tonal information but inverts the meaning of said name, sometimes asserting the opposite of what the name intends. This paper describes the problem, illustrates the inadequacy of current solutions, and proposes the adoption of the caron and circumflex marks. These are symbols with precedent in Yorùbá phonological scholarship since Olmsted (1951), used as orthographic conventions on single vowels to encode rising and falling contour tones, making them accessible for the first time through standard keyboard input and computational text processing. The proposal is supported by an implementation in the WriteYoruba keyboard and the TTSYoruba speech synthesizer, whose architecture and listener evaluation are reported separately (Tubosun et al., 2026).

Tue 29 SeptComputation and Language
The gist
Yorùbá is a language where the way the voice rises and falls in words matters for meaning. The authors found that current spelling methods often miss these patterns, sometimes changing the meaning of names unintentionally. They suggest using special marks called caron and circumflex on vowels to show these rising and falling tones clearly. This new way can be typed easily on keyboards and helps computers read Yorùbá better. The authors also made a keyboard and speech tool that use these marks.
Open → 2609.38627v1

Prosody-based speech features show no clear benefit beyond fusion method

Does a prosody-trained representation help beyond trainable fusion? A parameter-matched study with frozen HuBERT

Abstract: Explicit prosodic cues may help automatic speech recognition (ASR) of spontaneous speech, but auxiliary representations typically require additional trainable components, making it unclear whether gains come from the auxiliary information or the fusion mechanism. We address this using a frozen HuBERT backbone and a 64-dimensional representation trained to predict log F0, voicing, Delta log F0, log energy, and spectral tilt. We compare a frozen-backbone recognizer (Baseline), trainable fusion with zero auxiliary input (Null), and the same fusion supplied with the learned representation (Learned). Across Buckeye, Switchboard, and AMI IHM, Null reduces WER by 0.71-1.45 points over Baseline, whereas Learned differs from Null by +0.07, -0.09, and +0.00 points, with no significant differences. However, removing or mismatching the representation at inference increases Learned WER. Thus, Learned depends on the representation yet shows no measurable incremental WER benefit over the parameter-matched control.

Tue 29 SeptComputation and Language
The gist
The study looks at whether adding special speech features related to how people speak, like pitch and energy, helps improve automatic speech recognition beyond just using smarter ways to combine information. The authors compared three setups with and without these extra features, keeping everything else the same. They found that while using the fusion method alone improved accuracy, adding the special speech features made almost no difference. This means the benefit likely comes from how the systems combine data, not from these additional speech features themselves.
Open → 2609.36754v1

Framework improves multimodal depression detection with self-updating pipelines

MERID: Multimodal Exploration via Recursive Self-Improvement Agents for Major Depression Analysis

Abstract: Major depressive disorder (MDD) severely impacts daily activities and quality of life. Detecting MDD involves multimodal data, such as interview recordings and sensor measurements. This is particularly challenging, as these heterogeneous modalities often demand distinct, customized prediction pipelines. Existing efforts to address this challenge have explored both manually engineered multimodal architectures and agent-assisted pipeline development. Despite their progress, it remains challenging to autonomously revise pipelines based on experimental feedback and carry verified improvements forward into subsequent designs. To this end, we propose Multimodal Exploration via Recursive Self-Improvement Agents for Major Depression Analysis (MERID). The framework develops depression pipelines through experience-based recursive self-improvement (RSI). Grounded State Construction (GSC) grounds experience by aligning multimodal records with subject-level depression targets. Coupled Pipeline Exploration (CPE) jointly modifies representations, fusion, and predictors to build successor pipelines for classification and severity estimation. Evidence-Guided Evolution (EGE) guides revisions through feedback and verifies gains under uncertainty in small depression cohorts before inheritance. Extensive experiments on depression benchmarks show that MERID achieves the best results on multiple tasks compared with multimodal and agent-based baselines. Further analysis highlights the value of acoustic and linguistic cues for depression detection. Our code is available at https://github.com/DiscoAILab/MERID

Mon 28 SeptArtificial IntelligenceMachine LearningMultiagent Systems
The gist
Detecting major depression is hard because it uses many types of information, like speech and sensors, that need different ways to analyze them. The authors introduce MERID, a system that learns from its own past attempts to improve how it detects and measures depression. It automatically updates its analysis methods using feedback and checks improvements carefully before keeping them. Their tests show MERID works better than other methods on several depression detection tasks. They also found that sounds and language in speech are especially helpful for spotting depression.
Open → 2609.36235v1

Autoencoders learn separate timing and shape from complex functional data

Functional Autoencoders for Amplitude-Phase Representation Learning

Abstract: Functional data are intrinsically infinite-dimensional, and often exhibit phase variation, where corresponding events occur at different times across observations. Existing linear dimension reduction methods struggle with nonlinear amplitude variation, while functional autoencoders without an explicit warp entangle temporal misalignment with shape. We propose the Amplitude--Phase Functional Autoencoders (AP-FAE), an unsupervised framework for functional data that spans both univariate and multivariate cases, with emphasis on the multivariate setting, and factorizes the latent space into separate amplitude and phase embeddings derived from all channels. A smooth functional decoder reconstructs channel-specific amplitude functions in canonical time, and a shared monotone, endpoint-preserving warp captures phase variation. We prove a bound linking amplitude recovery to registration, reconstruction, and noise errors, and validate it numerically. Across synthetic data and six real-world benchmarks, AP-FAE outperforms state-of-the-art baselines on most clustering and alignment metrics and on all reconstruction metrics. Clustering with amplitude embeddings alone consistently surpasses joint amplitude--phase clustering, confirming the benefit of explicit disentanglement. Code is available at https://anonymous.4open.science/r/APFAE-418C/}{https://anonymous.4open.science/r/APFAE-418C/.

Mon 28 SeptMachine Learning
The gist
Functional data like heartbeats or motion can look very different if events happen at different times or with different intensities. The authors propose a new autoencoder method that can separate when things happen (phase) from how strong or shaped they are (amplitude) across multiple related signals. This separation helps the method better align and group complex signal data while reconstructing it accurately. Their method outperforms previous approaches on various tests and real-world datasets.
Open → 2609.34207v1

Fa-bench compares word and phone alignment in clean and noisy speech

FA-Bench: A Benchmark for Word-Level and Phone-Level Forced-Alignment and ASR Timestamps Under Clean and Noisy Conditions

Abstract: Forced alignment aligns speech audio with a text transcript to generate word and phone timestamps. Published comparisons normalize transcripts, split the data and match boundaries differently, so their numbers cannot be read together. We present FA-Bench, an open framework that fixes those choices once and releases the code, splits, phone mapping, text normalization and scoring script, with results published periodically. Track 1 gives every aligner the reference transcript and Track 2 gives it a recognizer's output, on the same audio, clean and degraded four ways, with 21 open models and 9 commercial APIs under a unified protocol. We score every boundary of an utterance and check the two labels beside it, so a word the recognizer missed or invented is charged. Using a tolerance-based F1 as our primary metric eliminates the 9% to 14% score inflation that standard MAE causes on recognition-dependent systems in conversational speech. We then group boundaries by their position and how many adjacent words were recognized correctly, which shows where a system lost the score. We discovered systematic bias in how current systems time words, with Whisper about 150 ms early and several commercial ASR APIs over 50 ms late. Code and results are at https://github.com/olewave/fa-bench

Sat 26 SeptComputation and LanguagePerformance
The gist
Forced alignment finds the timing of words and sounds in a spoken audio given a transcript, but different studies have used different methods that make it hard to compare results. The authors created FA-Bench, a shared tool and set of rules to test many alignment systems fairly on the same audio, both clean and noisy. They tested 21 open models and 9 commercial speech recognition services, showing how well each matches the transcript timing and where errors happen. They also found some systems consistently mistime words by a small but consistent delay or advance. FA-Bench's code and updated results are freely available.
Open → 2609.32396v1

Pretrained speech models improved for noisy police radio audio

Pretrained ASR Pseudo-labeling for Noisy Police Audio

Abstract: Pretrained ASR systems perform poorly on noisy Broadcast Police Communication (BPC), hindering efforts to understand police decision-making. Pseudo-labeling offers an unsupervised path to improve ASR without expensive human labels, but the efficacy of this approach on very noisy domains is not known. In this work, we systematically assess the opportunities and limits of pseudo-labeling to adapt foundation ASR models (Whisper and Qwen3-ASR) to noisy BPC domain corpora from Baltimore and Chicago. We demonstrate that existing internal confidence metrics (log-probabilities and STAR scores) fail to distinguish between high and low quality BPC pseudo-labels, and we introduce an external LLM-as-a-judge filtering paradigm that leverages parametric knowledge to discard contextually implausible transcripts. Our LLM-judging filters more aggressively than internal metrics and significantly reduces WER of the pseudo-labeled training sets across the Baltimore and Chicago BPC corpora, though a substantial gap remains relative to an oracle filter. We also introduce a new cross-model pseudo-labeling paradigm where one model is finetuned with pseudo-labels from the other, and we identify this method as a promising direction for future pseudo-labeling work.

Thu 24 SeptArtificial Intelligence
The gist
Speech recognition systems struggle to understand noisy police radio recordings, making it hard to analyze police decisions. The authors studied a way to improve these systems without expensive human labels by using machine-generated transcripts (pseudo-labels). They found that usual confidence scores cannot reliably pick good transcripts from bad ones and introduced a new method using large language models to filter out unlikely sentences. They also tested using pseudo-labels from one model to train another, showing this approach could help future improvements.
Open → 2609.30469v1

Speech enhancement model improves transcription accuracy on real recordings

Transcript-Supervised Post-Training of Generative Speech Enhancement on Real Recordings via Reinforce Adjoint Matching

Abstract: We adapt Reinforce Adjoint Matching (RAM), a reward-based post-training method, to generative speech enhancement (SE). Starting from a pretrained SE model, RAM tilts the model's conditional distribution toward outputs with higher reward. During training, the current model generates enhanced speech on-policy, evaluates each generated endpoint with a potentially non-differentiable reward, and analytically re-noises the endpoint to construct inputs for a reward-guided regression objective. This enables post-training directly on real recordings using weak supervision, such as text transcripts, without requiring paired clean speech targets or reward gradients. We investigate word error rate (WER)-based post-training and whether recognition performance can be improved without compromising perceptual speech quality. Experiments on real CHiME-4 recordings reduce WER by 5.08 percentage points relative to pretrained FlowSE without reducing any of the reported non-intrusive speech quality metrics. A subjective listening test at the default reward scale finds no statistically significant preference between the post-trained and pretrained models.

Thu 24 SeptMachine Learning
The gist
Speech enhancement systems aim to improve the clarity of speech recorded in noisy environments. The authors show how to better train these systems using real noisy recordings and just their transcriptions, without needing clean reference audio. They adapt a technique called Reinforce Adjoint Matching to guide the model toward reducing speech recognition errors. Their improved model lowers the word error rate on real recordings without making the enhanced speech sound worse.
Open → 2609.29405v1

Speech lLM pruning improved by component specific layer scoring

Speech Block Influence: Component-Specific Layer Scoring for Pruning Speech LLMs

Abstract: Speech LLMs are costly to deploy in resource-constrained settings. Layer pruning can cut this cost, but existing scoring metrics transfer poorly to speech LLMs: they assume a decoder-only architecture with homogeneous token sequences, whereas speech LLMs add encoder and adapter components and process multimodal sequences. We propose Speech Block Influence (SBI), the first layer-importance scoring framework designed for speech LLM pruning that consists of two component-specific scores: SBI-Enc measures the effect of encoder-layer removal at the adapter's output to better reflect downstream impact; SBI-Dec measures layer-wise input-output similarity over text-token positions only to avoid audio-token dominance. Across three speech LLMs, SBI improves pruning robustness, with stronger encoder performance at higher pruning rates and more reliable decoder layer selection by scoring text tokens rather than the audio-dominated full sequence. We further find that text-only calibration yields decoder rankings highly correlated with those from speech-text calibration, suggesting a cheaper alternative to measure decoder layer importance.

Thu 24 SeptSound
The gist
Speech large language models (LLMs) are often too big and slow for devices with limited resources. The authors found that existing ways to decide which parts of these models to cut don’t work well because speech LLMs process both audio and text in complicated ways. They created a new method called Speech Block Influence that scores different parts of the model separately based on their roles, making it easier and safer to trim the model without hurting its performance. Their method also shows that focusing only on text parts might be enough to figure out which layers are important, saving effort.
Open → 2609.29343v1

Compression of speech models increases fairness gaps for some groups

Temporal Taxation Compounds Under Post-Training Compression of Whisper Models

Abstract: 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.

Wed 23 SeptComputation and LanguageSound
The gist
Speech recognition models are often made smaller or faster before being used, but these changes can make them less fair for certain groups of people. The authors found that one type of model shrinking, called pruning, makes errors happen more often for some racial groups compared to others, meaning those people would need to spend more time fixing mistakes. Another shrinking method, called distillation, tends to reduce these fairness gaps. The study shows it’s important to check fairness after making these smaller models, not just before.
Open → 2609.28739v1

New method measures second language pronunciation using native speech data

A Native-Reference Coordinate Geometry for L2 Pronunciation Deviation Using Self-Supervised Speech Models

Abstract: Self-supervised speech models encode rich phonetic information, but it remains unclear how to transform this information into interpretable metrics for second-language (L2) pronunciation assessment in spontaneous speech. We propose a native-reference coordinate geometry in which phone-class averages from native speech define a low-dimensional reference subspace, and L2 speech is evaluated by its distance to matching native phone-class coordinates. Unlike prior distance-based approaches, our method does not require parallel recordings with matched linguistic content or dedicated pronunciation labels. Across different self-supervised encoders and modeling choices, the resulting native-reference distances show negative Spearman correlations up to -0.5 with speaking proficiency, indicating that higher-proficiency speakers tend to lie closer to the native-reference space.

Wed 23 SeptComputation and Language
The gist
It can be hard to tell how well someone speaks a second language just by listening. The authors created a way to compare a non-native speaker's pronunciation to native speakers using patterns learned by a computer from lots of speech. Their method does not need specially recorded matching sentences or labels. They found that better speakers tend to sound closer to native speech using this approach.
Open → 2609.28060v1

Hierarchical model improves word retrieval from non-invasive brain signals

HDND: Hierarchical Dynamic Neural Decoding for Multilingual Word/Character Retrieval from Non-Invasive Brain Recordings

Abstract: While deep learning has enabled language decoding from intracranial brain recordings, extending this capability to non-invasive recordings remains an unresolved challenge. Decoding individual words from non-invasive brain recordings is particularly difficult, as word-level neural evidence is weak, temporally distributed, and entangled with acoustic, lexical, and semantic structure. Existing retrieval pipelines often collapse these factors into a single representation, potentially discarding information available at intermediate temporal scales. Here, we introduce Hierarchical Dynamic Neural Decoding (HDND), a hierarchical dynamic decoding framework that treats word decoding as structured refinement rather than flat label retrieval. HDND combines intermediate neural representations, contextual semantic predictions, and, for selected reading conditions, an auxiliary character-form objective. We evaluate HDND across seven electroencephalography (EEG) and magnetoencephalography (MEG) datasets spanning English, Dutch, Mandarin, and Cantonese listening, reading, and reading-aloud conditions. Across the nine-condition word-retrieval benchmark, the proposed HDND yields a higher participant-averaged balanced Top-10 point estimate than the matched contextual word-decoding baseline in every condition and achieves the highest mean among all compared methods in eight of nine conditions. Across the same nine matched conditions, HDND also yields higher token-micro and pooled word-macro Top-10 point estimates in every setting. Sentence retrieval favors HDND in eight of nine conditions, while auditory speech-segment retrieval is mixed across the six listening conditions. These results show that hierarchical residual refinement can improve multilingual word retrieval from heterogeneous non-invasive brain recordings.

Mon 21 SeptComputer Vision and Pattern Recognition
The gist
Decoding individual words from non-invasive brain recordings is hard because the brain signals are weak and mixed with many other factors. The authors developed a new method called HDND that breaks down the decoding process into smaller steps, refining the word guess over time instead of guessing all at once. They tested HDND on brain signals recorded while people listened to, read, or spoke words in four different languages. This method consistently outperformed previous approaches in identifying the correct words from brain activity. The results suggest that using hierarchical, stepwise models can better capture the complex brain signals involved in language.
Open → 2609.24095v1