Towards Quantifying Benchmark Optimization in ASR Models

Public benchmarks are important measures of Automatic Speech Recognition (ASR) model capabilities. However, by nature of being public, there is risk of models being optimized for these benchmarks in ways that do not generalize well to real-world data. We present a methodology for quantifying benchmark optimization, focusing on cases where the audio underdetermines the reference transcript. We identify three families of behavioral probes that reveal models' capabilities of reproducing benchmark reference spans despite underdetermined audio: reference disagreement, masked-number recovery, and orthographic switching. We find that the highest-scoring open source models output verbatim reference transcript spans even when the relevant audio is contradictory, masked, or ambiguous. Using a variety of mechanistic probes, we show that models respond to narrow acoustic cues to override the faithful representation of the audio in favor of a benchmark-optimized policy. We show the benchmark-optimized behavior can be causally manipulated via low-rank linear steering or simply appending audio to the end of a segment in some cases. Overall, our results indicate that high-performing models exhibit benchmark-conditioned behaviors that can inflate benchmark performance without reflecting improved general-purpose transcription ability.

A Speech Corpus for Mizo Automatic Speech Recognition: Whisper and SraVaani 1.0 Fine-Tuning with Morphology-Aware Evaluation

This study reports the development of an Automatic Speech Recognition (ASR) system in Mizo, a low-resource language. The development included collecting 17.62 hours of speech data, curating it, and fine-tuning the Mizo ASR system with three Whisper multilingual models and with the SraVaani 1.0 Indic multilingual model. Whisper-large-v3 achieved the lowest conventional WER (18.08%), while morphology-aware evaluation yielded a WER of 7.22%. Zero-shot evaluation of the SraVaani 1.0 Indic multilingual model yielded a WER of 58.27%, while Mizo-specific fine-tuning reduced the conventional WER to 29.45% and the morphology-aware WER to 17.93%. The results demonstrate that the Whisper model can achieve a substantially low WER, even when adapted to an unseen language. In contrast, SraVaani 1.0 supports the Mizo language in its multilingual model; however, fine-tuning with carefully curated Mizo speech data substantially improves its performance.

Understanding Multilingual Medical ASR Adaptation Through Layer-Wise Analysis

Medical automatic speech recognition (MedASR) requires adaptation to specialised terminology, limited annotated clinical data, and multilingual use cases. Although large-scale pretrained ASR models such as Whisper achieve strong generalisation, their behaviour after medical and multilingual adaptation remains insufficiently understood beyond word error rate (WER). This paper investigates how multilingual medical adaptation reshapes the internal representations of Whisper models through layer-wise encoder analysis. We compare zero-shot decoding, English-only fine-tuning, German-only diagnostic fine-tuning, two-stage EN->EN+DE continuation, and direct EN+DE fine-tuning across Whisper model sizes. Fine-tuning substantially improves MedASR performance, but the best model depends on the adaptation setting: Whisper-Medium gives the lowest English WER (7.72%) and the lowest combined EN+DE WER under direct EN+DE training (26.30%); German-only Whisper-Large-v3 gives the lowest German WER (44.96%), but as a within-corpus diagnostic on 86 single-speaker training utterances rather than robust generalisation. Layer-wise analysis of the two-stage Whisper-Small trajectory shows that English medical fine-tuning produces the dominant encoder shift, whereas multilingual continuation largely preserves the adapted representation space. Domain and language information remain highly recoverable across layers, while linearly recoverable error-predictive cues weaken as WER improves.

SpeechSense: A Paralinguistic-Focused Dataset for Fine-Grained Speech Sentiment Analysis

Recent advances in AI have revolutionized speech processing, yet effective speech understanding requires discerning not just what is said, but how it is said. Speech Sentiment Analysis plays a critical role in decoding these paralinguistic cues for diverse real-world applications such as recruitment and customer service. However, existing Speech Sentiment Analysis research faces two primary limitations. First, dominant approaches rely on text-centric pipelines that cascade Automatic Speech Recognition with text analysis. This process inevitably discards essential acoustic features like prosody and tone, failing to capture attitudinal meanings in acoustically ambiguous utterances. Second, current benchmarks suffer from a mismatch in label granularity, prioritizing basic emotions (e.g., happy, sad) over the nuanced interpersonal stances (e.g., confident, impatient) necessary for social sensitivity. To address these limitations, we propose a novel dataset, SpeechSense, for fine-grained speech sentiment analysis. Specifically, we define a specialized 8-class taxonomy of interpersonal stances detectable primarily through prosodic cues beyond lexical content alone. We then construct a curated dataset based on this taxonomy, built from high-fidelity speech synthesis and rigorous human validation. Comprehensive experiments across multi-modal LLMs, text-only LLMs, and speech encoders demonstrate that models with acoustic access consistently outperform text-only baselines. These results empirically validate the primacy of acoustic cues in detecting subtle speaker attitudes, highlighting the necessity of SpeechSense. Dataset and supplementary materials are available at https://github.com/Sher13cked/SpeechSense.

Cached LLM Probability Retrieval for Speech Recognition

Large language models (LLMs) enhance automatic speech recognition (ASR) by providing linguistic priors; however, their direct rescoring is costly because it requires evaluating every N-best hypothesis. This paper introduces "cached LLM probability retrieval," which involves querying a local teacher LLM offline to obtain next-token probabilities for ASR-relevant context-target pairs. These probabilities are then utilized during recognition via cache lookups, backoff strategies, and optional scoring for significant misses. The method is training-free and can integrate with existing recognizers without requiring modifications to acoustic models. Evaluations across various ASR models reveal that cached retrieval outperforms 1-pass ASR in 28 of 39 settings and achieves lower non-oracle errors. Context length analysis indicates that benefits peak at a context length of 8, suggesting that cached probability retrieval is an effective and lightweight ASR adaptation method, in contrast to the heavy training required for Generative Error Correction (GER) or knowledge distillation (KD).

Aslema at NADI 2026: Augmentation through Fewshot for SLU

We present Aslema, our system for NADI 2026 Shared Task 5, which consists of two subtasks: intent recognition and slot filling. We evaluate four omni LLMs in a zero-shot setting and compare them with fine-tuned models. Our results show that fine-tuning consistently outperforms zero-shot inference. We further explore synthetic data augmentation by using an LLM to generate culturally grounded Tunisian Derja utterances, followed by voice cloning to generate synthetic speech. Incorporating this synthetic data improves performance on both tasks. Our final submitted system, based on Qwen3-Omni-30B and trained with a mixture of original and synthetic data, achieves 86.8% intent accuracy and 34.7 WER on the devtest split. On the official test set it ranks 1st in slot filling (59.5 CoER) and 4th among 8 teams in intent recognition (66.1% accuracy). We release our experimental scripts and will soon share the synthetic dataset to support further research in this area.

Emotion Across Speech and Faces: Shared Affective Mechanisms in Multimodal Foundation Models

Modern multimodal foundation models (MFMs) have made rapid progress on tasks requiring integrated perception across speech, vision, and language, including emotion recognition. However, it remains unclear whether they recognize speech and facial emotion through shared affective functional units or modality-specific pathways. We explore emotion-sensitive neurons (ESNs), sparse decoder neurons selectively associated with emotion categories, in three MFMs: Gemma-4-12B-it, MiniCPM-o-4.5, and Qwen2.5-Omni-7B. Using speech emotion recognition and facial expression recognition as complementary probes, we identify acoustic and visual ESNs. Visual ESNs are causally meaningful: deactivating them selectively impairs recognition of the associated facial emotion, whereas steering their activations selectively enhances recognition of that emotion relative to other emotion categories. Acoustic and visual ESNs further show emotion-matched overlap and similar layer-wise distributions, indicating partial structural alignment between affective representations across speech and faces. Finally, cross-modal interventions reveal bidirectional causal transfer: ESNs identified from one modality produce emotion-specific effects when applied to the other. Our findings provide one of the first cross-modality activation-level analyses of affective functional units in MFMs, suggesting that speech and facial emotion recognition partially converge onto sparse decoder-level components that can be localized and manipulated without training.

Navigating Speech Enhancement for Real-Time MRI: A Systematic Assessment of Signal Quality, Source Preservation, and Downstream Tasks

Audio recorded during real-time magnetic resonance imaging (rtMRI) is heavily contaminated by scanner noise, but it remains unclear whether general-purpose speech enhancement improves the signal for speech research and downstream processing. Three off-the-shelf systems---Denoiser, PASE, and RE-USE---are evaluated across five rtMRI corpora using naturally recorded inputs, a clean-input probe, and an archived paired additive-noise probe. The multi-task evaluation spans learned quality predictors, speaker and phone representations, reference-based intelligibility and quality measures, acoustic--phonetic probes, automatic speech recognition (ASR), and paralinguistic tasks. The central result is that enhancement effects are endpoint dependent: higher predicted-quality scores do not reliably imply better ASR performance or greater source fidelity. Across 15 corpus--recognizer comparisons using corpus-provided processed inputs, RE-USE yielded lower word-error-rate point estimates in 11, whereas Denoiser yielded higher estimates in 13. In the paired additive-noise probe, PASE and RE-USE improved recognized-phone agreement, intelligibility, and perceptual-quality point estimates. Denoiser improved recognized-phone agreement and short-time objective intelligibility (STOI) but reduced speaker-embedding similarity. No system was uniformly best across corpora, recognizers, and endpoints. Enhanced rtMRI audio should therefore be treated as a task-specific transformed derivative rather than a universally improved replacement for the original or DSP-processed waveform.

Unadapted Multilingual ASR on a Garrusi Kurdish Evaluation Set: A Common-Reference Staged Normalization Analysis

Evaluating speech recognition for a Kurdish variety written in a Latin field orthography, using a model that outputs Arabic script, creates a measurement problem before a modelling one: direct scoring treats writing-system differences as recognition errors. Jointly normalizing reference and hypothesis avoids this, but also changes reference tokenization, mixing agreement gains with a change in the scoring denominator. I evaluate MMS-1B-all with the Central Kurdish (ckb) adapter, used as released without adaptation, on 1,722 Garrusi questionnaire segments from five speakers (9,763 reference word tokens; 117.9 minutes). I use a common-reference design: the reference is folded once and fixed at 9,763 tokens, while only the hypothesis representation varies. The raw Arabic-script hypothesis scores 111.70% WER and 100.92% CER, with zero exact word matches. Latin transliteration gives 102.36% WER and 57.89% CER; folding it into the reference's reduced orthography gives 97.85% and 51.20%. Thus RAW-to-FOLDED reduces measured WER by 13.85 points and CER by 49.72 points; folding alone accounts for 4.51 and 6.69 points. Substantial error remains: 14.53% of reference tokens are exact matches, edits are substitution-dominated, and per-segment WER is higher for shorter segments. A Southern Kurdish fine-tuned system (aranemini/southern-kurdish-asr), scored under the same design, performs worse on every speaker (1,703 segments), with 109.56% WER and 55.85% CER. However, 12,330 output characters fall outside the folding table, so these rates must be recomputed against the corrected fixed reference. The MMS output also contains 613 unconverted or unmapped characters, showing that part of the residual error reflects scoring-pipeline limits rather than recognition alone. I will release the fixed reference and segment-level results, subject to source-corpus sharing terms, to support independent checking.

StreamHear: Domain-Adapted Pseudo-Labeling for Semi-Supervised Streaming Speech Recognition

Streaming automatic speech recognition (ASR) underperforms on domain-shifted target audio, where labeled in-domain data is costly to prepare while unlabeled audio is abundant. We present StreamHear, a semi-supervised pipeline that adapts a pretrained streaming student by fine-tuning an offline transducer teacher on the labeled training set, generating pseudo-labels on the unlabeled portion, and fine-tuning the student on the mixture. We further introduce a prior-regularized dynamic-programming realignment step that fixes chunk-level word placement using an ASR-hypothesis anchor. Across four datasets spanning financial calls, prepared read speech, and phone-quality dialogue, StreamHear consistently outperforms supervised student fine-tuning and narrows the gap to the offline teacher.