UniVoice: Unifying Autoregressive ASR and Flow-Matching based TTS with Large Language Models

Large language models (LLMs) have demonstrated promising performance in both automatic speech recognition (ASR) and text-to-speech (TTS) systems, gradually becoming the mainstream approach. However, most current approaches address these tasks separately rather than through a unified framework. This work aims to integrate these two tasks into one unified model. Although discrete speech tokenization enables joint modeling, its inherent information loss limits performance in both recognition and generation. In this work, we present UniVoice, a unified LLM framework through continuous representations that seamlessly integrates speech recognition and synthesis within a single model. Our approach combines the strengths of autoregressive modeling for speech recognition with flow matching for high-quality generation. To mitigate the inherent divergence between autoregressive and flow-matching models, we further design a dual attention mechanism, which switches between a causal mask for recognition and a bidirectional attention mask for synthesis. Furthermore, the proposed text-prefix-conditioned speech infilling method enables high-fidelity zero-shot voice cloning. Experimental results demonstrate that our method can achieve or exceed current single-task modeling methods in both ASR and zero-shot TTS tasks. This work explores new possibilities for end-to-end speech understanding and generation.

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.

Enhancing Speech Emotion Recognition via Fine-Tuning Pre-Trained Models and Hyper-Parameter Optimisation

We propose a workflow for speech emotion recognition (SER) that combines pre-trained representations with automated hyperparameter optimisation (HPO). Using SpeechBrain wav2vec2-base model fine-tuned on IEMOCAP as the encoder, we compare two HPO strategies, Gaussian Process Bayesian Optimisation (GP-BO) and Tree-structured Parzen Estimators (TPE), under an identical four-dimensional search space and 15-trial budget, with balanced class accuracy (BCA) on the German EmoDB corpus as the objective. All experiments run on 8 CPU cores with 32 GB RAM. GP-BO achieves 0.96 BCA in 11 minutes, and TPE (Hyperopt implementation) attains 0.97 in 15 minutes. In contrast, grid search requires 143 trials and 1,680 minutes to exceed 0.9 BCA, and the best AutoSpeech 2020 baseline reports only 0.85 in 30 minutes on GPU. For cross-lingual generalisation, an EmoDB-trained HPO-tuned model improves zero-shot accuracy by 0.25 on CREMA-D and 0.26 on RAVDESS. Results show that efficient HPO with pre-trained encoders delivers competitive SER on commodity CPUs. Source code to this work is available at: https://github.com/youngaryan/speechbrain-emotion-hpo.

Decoding Deception: Understanding Automatic Speech Recognition Vulnerabilities in Evasion and Poisoning Attacks

Recent studies have demonstrated the vulnerability of Automatic Speech Recognition systems to adversarial examples, which can deceive these systems into misinterpreting input speech commands. While previous research has primarily focused on white-box attacks with constrained optimizations, and transferability based black-box attacks against commercial Automatic Speech Recognition devices, this paper explores cost efficient white-box attack and non transferability black-box adversarial attacks on Automatic Speech Recognition systems, drawing insights from approaches such as Fast Gradient Sign Method and Zeroth-Order Optimization. Further, the novelty of the paper includes how poisoning attack can degrade the performances of state-of-the-art models leading to misinterpretation of audio signals. Through experimentation and analysis, we illustrate how hybrid models can generate subtle yet impactful adversarial examples with very little perturbation having Signal Noise Ratio of 35dB that can be generated within a minute. These vulnerabilities of state-of-the-art open source model have practical security implications, and emphasize the need for adversarial security.

MedVoiceBias: A Controlled Study of Audio LLM Behavior in Clinical Decision-Making

As large language models transition from text-based interfaces to audio interactions in clinical settings, they might introduce new vulnerabilities through paralinguistic cues in audio. We evaluated these models on 170 clinical cases, each synthesized into speech from 36 distinct voice profiles spanning variations in age, gender, and emotion. Our findings reveal a severe modality bias: surgical recommendations for audio inputs varied by as much as 35% compared to identical text-based inputs, with one model providing 80% fewer recommendations. Further analysis uncovered age disparities of up to 12% between young and elderly voices, which persisted in most models despite chain-of-thought prompting. While explicit reasoning successfully eliminated gender bias, the impact of emotion was not detected due to poor recognition performance. These results demonstrate that audio LLMs are susceptible to making clinical decisions based on a patient's voice characteristics rather than medical evidence, a flaw that risks perpetuating healthcare disparities. We conclude that bias-aware architectures are essential and urgently needed before the clinical deployment of these models.

Bloodroot: When Watermarking Turns Poisonous For Stealthy Backdoor

Backdoor data poisoning is a crucial technique for ownership protection and defending against malicious attacks. Embedding hidden triggers in training data can manipulate model outputs, enabling provenance verification, and deterring unauthorized use. However, current audio backdoor methods are suboptimal, as poisoned audio often exhibits degraded perceptual quality, which is noticeable to human listeners. This work explores the intrinsic stealthiness and effectiveness of audio watermarking in achieving successful poisoning. We propose a novel Watermark-as-Trigger concept, integrated into the Bloodroot backdoor framework via adversarial LoRA fine-tuning, which enhances perceptual quality while achieving a much higher trigger success rate and clean-sample accuracy. Experiments on speech recognition (SR) and speaker identification (SID) datasets show that watermark-based poisoning remains effective under acoustic filtering and model pruning. The proposed Bloodroot backdoor framework not only secures data-to-model ownership, but also well reveals the risk of adversarial misuse.

How I Built ASR for Endangered Languages with a Spoken Dictionary

Nearly half of the world's languages are endangered. Speech technologies such as Automatic Speech Recognition (ASR) are central to revival efforts, yet most languages remain unsupported because standard pipelines expect utterance-level supervised data. Speech data often exist for endangered languages but rarely match these formats. Manx Gaelic ($\sim$2,200 speakers), for example, has had transcribed speech since 1948, yet remains unsupported by modern systems. In this paper, we explore how little data, and in what form, is needed to build ASR for critically endangered languages. We show that a short-form pronunciation resource is a viable alternative, and that 40 minutes of such data produces usable ASR for Manx ($<$50\% WER). We replicate our approach, applying it to Cornish ($\sim$600 speakers), another critically endangered language. Results show that the barrier to entry, in quantity and form, is far lower than previously thought, giving hope to endangered language communities that cannot afford to meet the requirements arbitrarily imposed upon them.

Evaluating Self-Supervised Speech Models via Text-Based LLMS

Self-Supervised Learning (SSL) has gained traction for its ability to learn rich representations with low labeling costs, applicable across diverse downstream tasks. However, assessing the downstream-task performance remains challenging due to the cost of extra training and evaluation. Existing methods for task-agnostic evaluation also require extra training or hyperparameter tuning. We propose a novel evaluation metric using large language models (LLMs). By inputting discrete token sequences and minimal domain cues derived from SSL models into LLMs, we obtain the mean log-likelihood; these cues guide in-context learning, rendering the score more reliable without extra training or hyperparameter tuning. Experimental results show a correlation between LLM-based scores and automatic speech recognition task. Additionally, our findings reveal that LLMs not only functions as an SSL evaluation tools but also provides inference-time embeddings that are useful for speaker verification task.

How much speech data is necessary for ASR in African languages? An evaluation of data scaling in Kinyarwanda and Kikuyu

The development of Automatic Speech Recognition (ASR) systems for low-resource African languages remains challenging due to limited transcribed speech data. While recent advances in large multilingual models like OpenAI's Whisper offer promising pathways for low-resource ASR development, critical questions persist regarding practical deployment requirements. This paper addresses two fundamental concerns for practitioners: determining the minimum data volumes needed for viable performance and characterizing the primary failure modes that emerge in production systems. We evaluate Whisper's performance through comprehensive experiments on two Bantu languages: systematic data scaling analysis on Kinyarwanda using training sets from 1 to 1,400 hours, and detailed error characterization on Kikuyu using 270 hours of training data. Our scaling experiments demonstrate that practical ASR performance (WER < 13\%) becomes achievable with as little as 50 hours of training data, with substantial improvements continuing through 200 hours (WER < 10\%). Complementing these volume-focused findings, our error analysis reveals that data quality issues, particularly noisy ground truth transcriptions, account for 38.6\% of high-error cases, indicating that careful data curation is as critical as data volume for robust system performance. These results provide actionable benchmarks and deployment guidance for teams developing ASR systems across similar low-resource language contexts. We release accompanying and models see https://github.com/SunbirdAI/kinyarwanda-whisper-eval

Interpreting the Role of Visemes in Audio-Visual Speech Recognition

Audio-Visual Speech Recognition (AVSR) models have surpassed their audio-only counterparts in terms of performance. However, the interpretability of AVSR systems, particularly the role of the visual modality, remains under-explored. In this paper, we apply several interpretability techniques to examine how visemes are encoded in AV-HuBERT a state-of-the-art AVSR model. First, we use t-distributed Stochastic Neighbour Embedding (t-SNE) to visualize learned features, revealing natural clustering driven by visual cues, which is further refined by the presence of audio. Then, we employ probing to show how audio contributes to refining feature representations, particularly for visemes that are visually ambiguous or under-represented. Our findings shed light on the interplay between modalities in AVSR and could point to new strategies for leveraging visual information to improve AVSR performance.