Adapting Speech Foundation Models with Large Language Models for Unified Speech Recognition

Unified speech recognition aims to perform auditory, visual, and audiovisual speech recognition within a single model framework. While speech foundation models (SFMs) have demonstrated remarkable performance in auditory tasks, their adaptation to multimodal scenarios remains underexplored. This paper presents UASR-LLM, a novel framework that adapts frozen SFMs to unified VSR, ASR, and AVSR tasks by leveraging large language models (LLMs) as text decoders. Our approach introduces visual representations into multiple SFM layers through visual injection modules, enabling multimodal input processing and unified hidden representations. The augmented SFMs connect with decoder-only LLMs via a feed-forward adaptor, where concatenated representations and instruction prompts guide speech transcription. We implement a twostage training strategy: visual injection pretraining followed by speech recognition finetuning. SFM parameters remain frozen throughout training, with only visual injection modules optimized initially, and LLMs finetuned using LoRA parameters subsequently. Experimental results demonstrate superior performance over state-of-the-art baselines across VSR, ASR, and AVSR tasks under both clean and noisy conditions. Ablation studies confirm generalization across various SFMs and LLMs, validating the proposed training strategy.

Far from the Shallow: Brain-Predictive Reasoning Embedding through Residual Disentanglement

Understanding how the human brain progresses from processing simple linguistic inputs to performing high-level reasoning is a fundamental challenge in neuroscience. While modern large language models (LLMs) are increasingly used to model neural responses to language, their internal representations are highly "entangled," mixing information about lexicon, syntax, meaning, and reasoning. This entanglement biases conventional brain encoding analyses toward linguistically shallow features (e.g., lexicon and syntax), making it difficult to isolate the neural substrates of cognitively deeper processes. Here, we introduce a residual disentanglement method that computationally isolates these components. By first probing an LM to identify feature-specific layers, our method iteratively regresses out lower-level representations to produce four nearly orthogonal embeddings for lexicon, syntax, meaning, and, critically, reasoning. We used these disentangled embeddings to model intracranial (ECoG) brain recordings from neurosurgical patients listening to natural speech. We show that: 1) This isolated reasoning embedding exhibits unique predictive power, accounting for variance in neural activity not explained by other linguistic features and even extending to the recruitment of visual regions beyond classical language areas. 2) The neural signature for reasoning is temporally distinct, peaking later (~350-400ms) than signals related to lexicon, syntax, and meaning, consistent with its position atop a processing hierarchy. 3) Standard, non-disentangled LLM embeddings can be misleading, as their predictive success is primarily attributable to linguistically shallow features, masking the more subtle contributions of deeper cognitive processing.

EchoMind: An Interrelated Multi-level Benchmark for Evaluating Empathetic Speech Language Models

Speech Language Models (SLMs) have made significant progress in spoken language understanding. Yet it remains unclear whether they can fully perceive non lexical vocal cues alongside spoken words, and respond with empathy that aligns with both emotional and contextual factors. Existing benchmarks typically evaluate linguistic, acoustic, reasoning, or dialogue abilities in isolation, overlooking the integration of these skills that is crucial for human-like, emotionally intelligent conversation. We present EchoMind, the first interrelated, multi-level benchmark that simulates the cognitive process of empathetic dialogue through sequential, context-linked tasks: spoken-content understanding, vocal-cue perception, integrated reasoning, and response generation. All tasks share identical and semantically neutral scripts that are free of explicit emotional or contextual cues, and controlled variations in vocal style are used to test the effect of delivery independent of the transcript. EchoMind is grounded in an empathy-oriented framework spanning 3 coarse and 12 fine-grained dimensions, encompassing 39 vocal attributes, and evaluated using both objective and subjective metrics. Testing 12 advanced SLMs reveals that even state-of-the-art models struggle with high-expressive vocal cues, limiting empathetic response quality. Analyses of prompt strength, speech source, and ideal vocal cue recognition reveal persistent weaknesses in instruction-following, resilience to natural speech variability, and effective use of vocal cues for empathy. These results underscore the need for SLMs that integrate linguistic content with diverse vocal cues to achieve truly empathetic conversational ability.

LRW-Persian: Lip-reading in the Wild Dataset for Persian Language

Lipreading has emerged as an increasingly important research area for developing robust speech recognition systems and assistive technologies for the hearing-impaired. However, non-English resources for visual speech recognition remain limited. We introduce LRW-Persian, the largest in-the-wild Persian word-level lipreading dataset, comprising $743$ target words and over $414{,}000$ video samples extracted from more than $1{,}900$ hours of footage across $67$ television programs. Designed as a benchmark-ready resource, LRW-Persian provides speaker-disjoint training and test splits, wide regional and dialectal coverage, and rich per-clip metadata including head pose, age, and gender. To ensure large-scale data quality, we establish a fully automated end-to-end curation pipeline encompassing transcription based on Automatic Speech Recognition(ASR), active-speaker localization, quality filtering, and pose/mask screening. We further fine-tune two widely used lipreading architectures on LRW-Persian, establishing reference performance and demonstrating the difficulty of Persian visual speech recognition. By filling a critical gap in low-resource languages, LRW-Persian enables rigorous benchmarking, supports cross-lingual transfer, and provides a foundation for advancing multimodal speech research in underrepresented linguistic contexts. The dataset is publicly available at: https://lrw-persian.vercel.app.

Look and Tell: A Dataset for Multimodal Grounding Across Egocentric and Exocentric Views

We introduce Look and Tell, a multimodal dataset for studying referential communication across egocentric and exocentric perspectives. Using Meta Project Aria smart glasses and stationary cameras, we recorded synchronized gaze, speech, and video as 25 participants instructed a partner to identify ingredients in a kitchen. Combined with 3D scene reconstructions, this setup provides a benchmark for evaluating how different spatial representations (2D vs. 3D; ego vs. exo) affect multimodal grounding. The dataset contains 3.67 hours of recordings, including 2,707 richly annotated referential expressions, and is designed to advance the development of embodied agents that can understand and engage in situated dialogue.

HyBeam: Hybrid Microphone-Beamforming Array-Agnostic Speech Enhancement for Wearables

Speech enhancement is a fundamental challenge in signal processing, particularly when robustness is required across diverse acoustic conditions and microphone setups. Deep learning methods have been successful for speech enhancement, but often assume fixed array geometries, limiting their use in mobile, embedded, and wearable devices. Existing array-agnostic approaches typically rely on either raw microphone signals or beamformer outputs, but both have drawbacks under changing geometries. We introduce HyBeam, a hybrid framework that uses raw microphone signals at low frequencies and beamformer signals at higher frequencies, exploiting their complementary strengths while remaining highly array-agnostic. Simulations across diverse rooms and wearable array configurations demonstrate that HyBeam consistently surpasses microphone-only and beamformer-only baselines in PESQ, STOI, and SI-SDR. A bandwise analysis shows that the hybrid approach leverages beamformer directivity at high frequencies and microphone cues at low frequencies, outperforming either method alone across all bands.

Mitigating Attention Sinks and Massive Activations in Audio-Visual Speech Recognition with LLMS

Large language models (LLMs) have recently advanced auditory speech recognition (ASR), visual speech recognition (VSR), and audio-visual speech recognition (AVSR). However, understanding of their internal dynamics under fine-tuning remains limited. In natural language processing, recent work has revealed attention sinks, tokens that attract disproportionately high attention, and associated massive activations in which some features of sink tokens exhibit huge activation in LLMs. In this work, we are the first to study these phenomena in multimodal speech recognition. Through a detailed analysis of audio-visual LLMs, we identify attention sinks and massive activations not only at the BOS token but also at intermediate low-semantic tokens across ASR, VSR, and AVSR. We show that massive activations originate in the MLP layers and correspond to fixed feature indices across all sink tokens. We further show that intermediate sink tokens exhibit high cosine similarity to the BOS token, thereby amplifying attention and activation. Building on these insights, we introduce a simple decorrelation loss that reduces cosine similarity between BOS and other tokens, effectively mitigating intermediate sinks and massive activations. Furthermore, our method improves word error rate (WER) under high audio-visual feature downsampling while remaining stable at lower downsampling rates.

UltraVoice: Scaling Fine-Grained Style-Controlled Speech Conversations for Spoken Dialogue Models

Spoken dialogue models currently lack the ability for fine-grained speech style control, a critical capability for human-like interaction that is often overlooked in favor of purely functional capabilities like reasoning and question answering. To address this limitation, we introduce UltraVoice, the first large-scale speech dialogue dataset engineered for multiple fine-grained speech style control. Encompassing over 830 hours of speech dialogues, UltraVoice provides instructions across six key speech stylistic dimensions: emotion, speed, volume, accent, language, and composite styles. Fine-tuning leading models such as SLAM-Omni and VocalNet on UltraVoice significantly enhances their fine-grained speech stylistic controllability without degrading core conversational abilities. Specifically, our fine-tuned models achieve improvements of 29.12-42.33% in Mean Opinion Score (MOS) and 14.61-40.09 percentage points in Instruction Following Rate (IFR) on multi-dimensional control tasks designed in the UltraVoice. Moreover, on the URO-Bench benchmark, our fine-tuned models demonstrate substantial gains in core understanding, reasoning, and conversational abilities, with average improvements of +10.84% on the Basic setting and +7.87% on the Pro setting. Furthermore, the dataset's utility extends to training controllable Text-to-Speech (TTS) models, underscoring its high quality and broad applicability for expressive speech synthesis. The complete dataset and model checkpoints are available at: https://github.com/bigai-nlco/UltraVoice.

A Sociophonetic Analysis of Racial Bias in Commercial ASR Systems Using the Pacific Northwest English Corpus

This paper presents a systematic evaluation of racial bias in four major commercial automatic speech recognition (ASR) systems using the Pacific Northwest English (PNWE) corpus. We analyze transcription accuracy across speakers from four ethnic backgrounds (African American, Caucasian American, ChicanX, and Yakama) and examine how sociophonetic variation contributes to differential system performance. We introduce a heuristically-determined Phonetic Error Rate (PER) metric that links recognition errors to specific linguistically motivated variables derived from sociophonetic annotation. Our analysis of eleven sociophonetic features reveals that vowel quality variation, particularly resistance to the low-back merger and pre-nasal merger patterns, is systematically associated with differential error rates across ethnic groups, with the most pronounced effects for African American speakers across all evaluated systems. These findings demonstrate that acoustic modeling of dialectal phonetic variation, rather than lexical or syntactic factors, remains a primary source of bias in commercial ASR systems. The study establishes the PNWE corpus as a valuable resource for bias evaluation in speech technologies and provides actionable guidance for improving ASR performance through targeted representation of sociophonetic diversity in training data.

The Limits of Data Scaling: Sub-token Utilization and Acoustic Saturation in Multilingual ASR

How much audio is needed to fully observe a multilingual ASR model's learned sub-token inventory across languages, and does data disparity in multilingual pre-training affect how these tokens are utilized during inference? We address this question by analyzing Whisper's decoding behavior during inference across 49 languages. By logging decoding candidate sub-tokens and tracking their cumulative discovery over time, we study the utilization pattern of the model's sub-token space. Results show that the total number of discovered tokens remains largely independent of a language's pre-training hours, indicating that data disparity does not strongly influence lexical diversity in the model's hypothesis space. Sub-token discovery rates follow a consistent exponential saturation pattern across languages, suggesting a stable time window after which additional audio yields minimal new sub-token activation. We refer to this convergence threshold as acoustic saturation time (AST). Further analyses of rank-frequency distributions reveal Zipf-like patterns better modeled by a Zipf-Mandelbrot law, and mean sub-token length shows a positive correlation with resource level. Additionally, those metrics show more favorable patterns for languages in the Latin script than those in scripts such as Cyrillic, CJK, and Semitic. Together, our study suggests that sub-token utilization during multilingual ASR inference is constrained more by the statistical, typological, and orthographic structure of the speech than by training data scale, providing an empirical basis for more equitable corpus construction and cross-lingual evaluation.