The goal of voice anonymization is to modify an audio such that the true identity of its speaker is hidden. Research on this task is typically limited to the same English read speech datasets, thus the efficacy of current methods for other types of speech data remains unknown. In this paper, we present the first investigation of voice anonymization for the multilingual phenomenon of code-switching speech. We prepare two corpora for this task and propose adaptations to a multilingual anonymization model to make it applicable for code-switching speech. By testing the anonymization performance of this and two language-independent methods on the datasets, we find that only the multilingual system performs well in terms of privacy and utility preservation. Furthermore, we observe challenges in performing utility evaluations on this data because of its spontaneous character and the limited code-switching support by the multilingual speech recognition model.
With the advent of new sequence models like Mamba and xLSTM, several studies have shown that these models match or outperform state-of-the-art models in single-channel speech enhancement, automatic speech recognition, and self-supervised audio representation learning. However, prior research has demonstrated that sequence models like LSTM and Mamba tend to overfit to the training set. To address this issue, previous works have shown that adding self-attention to LSTMs substantially improves generalization performance for single-channel speech enhancement. Nevertheless, neither the concept of hybrid Mamba and time-frequency attention models nor their generalization performance have been explored for speech enhancement. In this paper, we propose a novel hybrid architecture, MambAttention, which combines Mamba and shared time- and frequency-multi-head attention modules for generalizable single-channel speech enhancement. To train our model, we introduce VoiceBank+Demand Extended (VB-DemandEx), a dataset inspired by VoiceBank+Demand but with more challenging noise types and lower signal-to-noise ratios. Trained on VB-DemandEx, our proposed MambAttention model significantly outperforms existing state-of-the-art LSTM-, xLSTM-, Mamba-, and Conformer-based systems of similar complexity across all reported metrics on two out-of-domain datasets: DNS 2020 and EARS-WHAM_v2, while matching their performance on the in-domain dataset VB-DemandEx. Ablation studies highlight the role of weight sharing between the time- and frequency-multi-head attention modules for generalization performance. Finally, we explore integrating the shared time- and frequency-multi-head attention modules with LSTM and xLSTM, which yields a notable performance improvement on the out-of-domain datasets. However, our MambAttention model remains superior on both out-of-domain datasets across all reported evaluation metrics.
The ability to dynamically adjust the computational load of neural models during inference in a resource aware manner is crucial for on-device processing scenarios, characterised by limited and time-varying computational resources. Early-exit architectures represent an elegant and effective solution, since they can process the input with a subset of their layers, exiting at intermediate branches (the upmost layers are hence removed from the model). From a different perspective, for automatic speech recognition applications there are memory-efficient neural architectures that apply variable frame rate analysis, through downsampling/upsampling operations in the middle layers, reducing the overall number of operations and improving significantly the performance on well established benchmarks. One example is the Zipformer. However, these architectures lack the modularity necessary to inject early-exit branches. With the aim of improving the performance in early-exit models, we propose introducing parallel layers in the architecture that process downsampled versions of their inputs. % in conjunction with standard processing layers. We show that in this way the speech recognition performance on standard benchmarks significantly improve, at the cost of a small increase in the overall number of model parameters but without affecting the inference time.
Small-Footprint Keyword Spotting (SF-KWS) has gained popularity in today's landscape of smart voice-activated devices, smartphones, and Internet of Things (IoT) applications. This surge is attributed to the advancements in Deep Learning, enabling the identification of predefined words or keywords from a continuous stream of words. To implement the SF-KWS model on edge devices with low power and limited memory in real-world scenarios, a efficient Tiny Machine Learning (TinyML) framework is essential. In this study, we explore seven distinct categories of techniques namely, Model Architecture, Learning Techniques, Model Compression, Attention Awareness Architecture, Feature Optimization, Neural Network Search, and Hybrid Approaches, which are suitable for developing an SF-KWS system. This comprehensive overview will serve as a valuable resource for those looking to understand, utilize, or contribute to the field of SF-KWS. The analysis conducted in this work enables the identification of numerous potential research directions, encompassing insights from automatic speech recognition research and those specifically pertinent to the realm of spoken SF-KWS.
Automatic Speech Recognition (ASR) has transformed daily tasks from video transcription to workplace hiring. ASR systems' growing use warrants robust and standardized auditing approaches to ensure automated transcriptions of high and equitable quality. This is especially critical for people with speech and language disorders (such as aphasia) who may disproportionately depend on ASR systems to navigate everyday life. In this work, we identify three pitfalls in existing standard ASR auditing procedures, and demonstrate how addressing them impacts audit results via a case study of six popular ASR systems' performance for aphasia speakers. First, audits often adhere to a single method of text standardization during data pre-processing, which (a) masks variability in ASR performance from applying different standardization methods, and (b) may not be consistent with how users - especially those from marginalized speech communities - would want their transcriptions to be standardized. Second, audits often display high-level demographic findings without further considering performance disparities among (a) more nuanced demographic subgroups, and (b) relevant covariates capturing acoustic information from the input audio. Third, audits often rely on a single gold-standard metric -- the Word Error Rate -- which does not fully capture the extent of errors arising from generative AI models, such as transcription hallucinations. We propose a more holistic auditing framework that accounts for these three pitfalls, and exemplify its results in our case study, finding consistently worse ASR performance for aphasia speakers relative to a control group. We call on practitioners to implement these robust ASR auditing practices that remain flexible to the rapidly changing ASR landscape.
Multi-state spiking neurons such as the adaptive leaky integrate-and-fire (AdLIF) neuron offer compelling alternatives to conventional deep learning models thanks to their sparse binary activations, second-order nonlinear recurrent dynamics, and efficient hardware realizations. However, such internal dynamics can cause instabilities during inference and training, often limiting performance and scalability. Meanwhile, state space models (SSMs) excel in long sequence processing using linear state-intrinsic recurrence resembling spiking neurons' subthreshold regime. Here, we establish a mathematical bridge between SSMs and second-order spiking neuron models. Based on structure and parametrization strategies of diagonal SSMs, we propose two novel spiking neuron models. The first extends the AdLIF neuron through timestep training and logarithmic reparametrization to facilitate training and improve final performance. The second additionally brings initialization and structure from complex-state SSMs, broadening the dynamical regime to oscillatory dynamics. Together, our two models achieve beyond or near state-of-the-art (SOTA) performances for reset-based spiking neuron models across both event-based and raw audio speech recognition datasets. We achieve this with a favorable number of parameters and required dynamic memory while maintaining high activity sparsity. Our models demonstrate enhanced scalability in network size and strike a favorable balance between performance and efficiency with respect to SSM models.
Automatic speech recognition (ASR) systems struggle with dysarthric speech due to high inter-speaker variability and slow speaking rates. To address this, we explore dysarthric-to-healthy speech conversion for improved ASR performance. Our approach extends the Rhythm and Voice (RnV) conversion framework by introducing a syllable-based rhythm modeling method suited for dysarthric speech. We assess its impact on ASR by training LF-MMI models and fine-tuning Whisper on converted speech. Experiments on the Torgo corpus reveal that LF-MMI achieves significant word error rate reductions, especially for more severe cases of dysarthria, while fine-tuning Whisper on converted data has minimal effect on its performance. These results highlight the potential of unsupervised rhythm and voice conversion for dysarthric ASR. Code available at: https://github.com/idiap/RnV
Self-Supervised Learning (SSL) has demonstrated strong performance in speech processing, particularly in automatic speech recognition. In this paper, we explore an SSL pretraining framework that leverages masked language modeling with targets derived from a speech recognition model. We also present chunkwise attention with dynamic chunk size sampling during pretraining to enable both full-context and streaming fine-tuning. Our experiments examine scaling with respect to model size and the amount of data. Using our method, we train the GigaAM family of models, including a state-of-the-art model for Russian speech recognition that outperforms Whisper-large-v3 by 50%. We have released our foundation and ASR models, along with the inference code, under the MIT license as open-source resources to the research community. Available at https://github.com/salute-developers/gigaam.
How language-specific are speech representations learned by self-supervised models? Existing work has shown that a range of linguistic features can be successfully decoded from end-to-end models trained only on speech recordings. However, it's less clear to what extent pre-training on specific languages improves language-specific linguistic information. Here we test the encoding of Dutch phonetic and lexical information in internal representations of self-supervised Wav2Vec2 models. Pre-training exclusively on Dutch improves the representation of Dutch linguistic features as compared to pre-training on similar amounts of English or larger amounts of multilingual data. This language-specific advantage is well-detected by trained clustering or classification probes, and partially observable using zero-shot metrics. Furthermore, the language-specific benefit on linguistic feature encoding aligns with downstream performance on Automatic Speech Recognition.
Sanskrit, an ancient language with a rich linguistic heritage, presents unique challenges for automatic speech recognition (ASR) due to its phonemic complexity and the phonetic transformations that occur at word junctures, similar to the connected speech found in natural conversations. Due to these complexities, there has been limited exploration of ASR in Sanskrit, particularly in the context of its poetic verses, which are characterized by intricate prosodic and rhythmic patterns. This gap in research raises the question: How can we develop an effective ASR system for Sanskrit, particularly one that captures the nuanced features of its poetic form? In this study, we introduce Vedavani, the first comprehensive ASR study focused on Sanskrit Vedic poetry. We present a 54-hour Sanskrit ASR dataset, consisting of 30,779 labelled audio samples from the Rig Veda and Atharva Veda. This dataset captures the precise prosodic and rhythmic features that define the language. We also benchmark the dataset on various state-of-the-art multilingual speech models.$^{1}$ Experimentation revealed that IndicWhisper performed the best among the SOTA models.
Dysarthria, a motor speech disorder, affects intelligibility and requires targeted interventions for effective communication. In this work, we investigate automated mispronunciation feedback by collecting a dysarthric speech dataset from six speakers reading two passages, annotated by a speech therapist with temporal markers and mispronunciation descriptions. We design a three-stage framework for explainable mispronunciation evaluation: (1) overall clarity scoring, (2) mispronunciation localization, and (3) mispronunciation type classification. We systematically analyze pretrained Automatic Speech Recognition (ASR) models in each stage, assessing their effectiveness in dysarthric speech evaluation (Code available at: https://github.com/augmented-human-lab/interspeech25_speechtherapy, Supplementary webpage: https://apps.ahlab.org/interspeech25_speechtherapy/). Our findings offer clinically relevant insights for automating actionable feedback for pronunciation assessment, which could enable independent practice for patients and help therapists deliver more effective interventions.
This paper discusses the construction, fine-tuning, and deployment of BeaverTalk, a cascaded system for speech-to-text translation as part of the IWSLT 2025 simultaneous translation task. The system architecture employs a VAD segmenter for breaking a speech stream into segments, Whisper Large V2 for automatic speech recognition (ASR), and Gemma 3 12B for simultaneous translation. Regarding the simultaneous translation LLM, it is fine-tuned via low-rank adaptors (LoRAs) for a conversational prompting strategy that leverages a single prior-sentence memory bank from the source language as context. The cascaded system participated in the English$\rightarrow$German and English$\rightarrow$Chinese language directions for both the low and high latency regimes. In particular, on the English$\rightarrow$German task, the system achieves a BLEU of 24.64 and 27.83 at a StreamLAAL of 1837.86 and 3343.73, respectively. Then, on the English$\rightarrow$Chinese task, the system achieves a BLEU of 34.07 and 37.23 at a StreamLAAL of 2216.99 and 3521.35, respectively.
Document-level context is crucial for handling discourse challenges in text-to-text document-level machine translation (MT). Despite the increased discourse challenges introduced by noise from automatic speech recognition (ASR), the integration of document-level context in speech translation (ST) remains insufficiently explored. In this paper, we develop DoCIA, an online framework that enhances ST performance by incorporating document-level context. DoCIA decomposes the ST pipeline into four stages. Document-level context is integrated into the ASR refinement, MT, and MT refinement stages through auxiliary LLM (large language model)-based modules. Furthermore, DoCIA leverages document-level information in a multi-level manner while minimizing computational overhead. Additionally, a simple yet effective determination mechanism is introduced to prevent hallucinations from excessive refinement, ensuring the reliability of the final results. Experimental results show that DoCIA significantly outperforms traditional ST baselines in both sentence and discourse metrics across four LLMs, demonstrating its effectiveness in improving ST performance.
Automatic Speech Recognition (ASR) systems struggle with child speech due to its distinct acoustic and linguistic variability and limited availability of child speech datasets, leading to high transcription error rates. While ASR error correction (AEC) methods have improved adult speech transcription, their effectiveness on child speech remains largely unexplored. To address this, we introduce CHSER, a Generative Speech Error Correction (GenSEC) dataset for child speech, comprising 200K hypothesis-transcription pairs spanning diverse age groups and speaking styles. Results demonstrate that fine-tuning on the CHSER dataset achieves up to a 28.5% relative WER reduction in a zero-shot setting and a 13.3% reduction when applied to fine-tuned ASR systems. Additionally, our error analysis reveals that while GenSEC improves substitution and deletion errors, it struggles with insertions and child-specific disfluencies. These findings highlight the potential of GenSEC for improving child ASR.
Despite the remarkable progress in end-to-end Automatic Speech Recognition (ASR) engines, accurately transcribing dysarthric speech remains a major challenge. In this work, we proposed a two-stage framework for the Speech Accessibility Project Challenge at INTERSPEECH 2025, which combines cutting-edge speech recognition models with LLM-based generative error correction (GER). We assess different configurations of model scales and training strategies, incorporating specific hypothesis selection to improve transcription accuracy. Experiments on the Speech Accessibility Project dataset demonstrate the strength of our approach on structured and spontaneous speech, while highlighting challenges in single-word recognition. Through comprehensive analysis, we provide insights into the complementary roles of acoustic and linguistic modeling in dysarthric speech recognition
In our prior works, we introduced a scalable streaming speech synthesis model, CosyVoice 2, which integrates a large language model (LLM) and a chunk-aware flow matching (FM) model, and achieves low-latency bi-streaming speech synthesis and human-parity quality. Despite these advancements, CosyVoice 2 exhibits limitations in language coverage, domain diversity, data volume, text formats, and post-training techniques. In this paper, we present CosyVoice 3, an improved model designed for zero-shot multilingual speech synthesis in the wild, surpassing its predecessor in content consistency, speaker similarity, and prosody naturalness. Key features of CosyVoice 3 include: 1) A novel speech tokenizer to improve prosody naturalness, developed via supervised multi-task training, including automatic speech recognition, speech emotion recognition, language identification, audio event detection, and speaker analysis. 2) A new differentiable reward model for post-training applicable not only to CosyVoice 3 but also to other LLM-based speech synthesis models. 3) Dataset Size Scaling: Training data is expanded from ten thousand hours to one million hours, encompassing 9 languages and 18 Chinese dialects across various domains and text formats. 4) Model Size Scaling: Model parameters are increased from 0.5 billion to 1.5 billion, resulting in enhanced performance on our multilingual benchmark due to the larger model capacity. These advancements contribute significantly to the progress of speech synthesis in the wild. We encourage readers to listen to the demo at https://funaudiollm.github.io/cosyvoice3.
The rapid growth of voice assistants powered by large language models (LLM) has highlighted a need for speech instruction data to train these systems. Despite the abundance of speech recognition data, there is a notable scarcity of speech instruction data, which is essential for fine-tuning models to understand and execute spoken commands. Generating high-quality synthetic speech requires a good text-to-speech (TTS) model, which may not be available to low resource languages. Our novel approach addresses this challenge by halting synthesis at the semantic representation level, bypassing the need for TTS. We achieve this by aligning synthetic semantic representations with the pre-trained Whisper encoder, enabling an LLM to be fine-tuned on text instructions while maintaining the ability to understand spoken instructions during inference. This simplified training process is a promising approach to building voice assistant for low-resource languages.
Generative error correction (GER) with large language models (LLMs) has emerged as an effective post-processing approach to improve automatic speech recognition (ASR) performance. However, it often struggles with rare or domain-specific words due to limited training data. Furthermore, existing LLM-based GER approaches primarily rely on textual information, neglecting phonetic cues, which leads to over-correction. To address these issues, we propose a novel LLM-based GER approach that targets rare words and incorporates phonetic information. First, we generate synthetic data to contain rare words for fine-tuning the GER model. Second, we integrate ASR's N-best hypotheses along with phonetic context to mitigate over-correction. Experimental results show that our method not only improves the correction of rare words but also reduces the WER and CER across both English and Japanese datasets.
Recent Multimodal Large Language Models (MLLMs) achieve promising performance on visual and audio benchmarks independently. However, the ability of these models to process cross-modal information synchronously remains largely unexplored. In this paper, we introduce: 1) Daily-Omni, an Audio-Visual Questioning and Answering benchmark comprising 684 videos of daily life scenarios from diverse sources, rich in both audio and visual information, and featuring 1197 multiple-choice QA pairs across 6 major tasks; 2) Daily-Omni QA Generation Pipeline, which includes automatic annotation, QA generation and QA optimization, significantly improves efficiency for human evaluation and scalability of the benchmark; 3) Daily-Omni-Agent, a training-free agent utilizing open-source Visual Language Model (VLM), Audio Language Model (ALM) and Automatic Speech Recognition (ASR) model to establish a baseline for this benchmark. The results show that current MLLMs still struggle significantly with tasks requiring audio-visual integration, but combining VLMs and ALMs with simple temporal alignment techniques can achieve substantially better performance. Codes and benchmark are available at \href{https://github.com/Lliar-liar/Daily-Omni}{https://github.com/Lliar-liar/Daily-Omni}.
Recent advances in Automatic Speech Recognition (ASR) have been largely fueled by massive speech corpora. However, extending coverage to diverse languages with limited resources remains a formidable challenge. This paper introduces Speech Back-Translation, a scalable pipeline that improves multilingual ASR models by converting large-scale text corpora into synthetic speech via off-the-shelf text-to-speech (TTS) models. We demonstrate that just tens of hours of real transcribed speech can effectively train TTS models to generate synthetic speech at hundreds of times the original volume while maintaining high quality. To evaluate synthetic speech quality, we develop an intelligibility-based assessment framework and establish clear thresholds for when synthetic data benefits ASR training. Using Speech Back-Translation, we generate more than 500,000 hours of synthetic speech in ten languages and continue pre-training Whisper-large-v3, achieving average transcription error reductions of over 30\%. These results highlight the scalability and effectiveness of Speech Back-Translation for enhancing multilingual ASR systems.