Mirco Ravanelli

Jama Hussein Mohamud, Mohsin Hasan, Mirco Ravanelli et al.

Masked diffusion models have seen great success in capturing data distributions over discrete sequences in domains such as text and proteins. These models generate data by iteratively unmasking tokens starting from a fully masked sequence, with the unmasking order typically chosen at random or using a heuristic based on denoiser probabilities. In this work, we propose a scheme for learning the unmasking order using an additional lightweight policy network on top of a diffusion model. Our proposed loss reweights terms in the masked diffusion loss according to policy probabilities, and results in a policy that prefers positions where the denoiser is more likely to be correct. We study this loss in two settings: (i) training solely the policy while using a frozen pre-trained denoiser, and (ii) training the policy and denoiser jointly with the weighted loss to allow for mutual adaptation. We demonstrate that our approach outperforms common heuristics on problems that are sensitive to token ordering, such as combinatorial tasks and proteins.

Luca Della Libera, Cem Subakan, Mirco Ravanelli et al. · cmu

Transformers have recently achieved state-of-the-art performance in speech separation. These models, however, are computationally demanding and require a lot of learnable parameters. This paper explores Transformer-based speech separation with a reduced computational cost. Our main contribution is the development of the Resource-Efficient Separation Transformer (RE-SepFormer), a self-attention-based architecture that reduces the computational burden in two ways. First, it uses non-overlapping blocks in the latent space. Second, it operates on compact latent summaries calculated from each chunk. The RE-SepFormer reaches a competitive performance on the popular WSJ0-2Mix and WHAM! datasets in both causal and non-causal settings. Remarkably, it scales significantly better than the previous Transformer-based architectures in terms of memory and inference time, making it more suitable for processing long mixtures.

Yingzhi Wang, Mirco Ravanelli, Alya Yacoubi

Speech Emotion Recognition (SER) typically relies on utterance-level solutions. However, emotions conveyed through speech should be considered as discrete speech events with definite temporal boundaries, rather than attributes of the entire utterance. To reflect the fine-grained nature of speech emotions, we propose a new task: Speech Emotion Diarization (SED). Just as Speaker Diarization answers the question of "Who speaks when?", Speech Emotion Diarization answers the question of "Which emotion appears when?". To facilitate the evaluation of the performance and establish a common benchmark for researchers, we introduce the Zaion Emotion Dataset (ZED), an openly accessible speech emotion dataset that includes non-acted emotions recorded in real-life conditions, along with manually-annotated boundaries of emotion segments within the utterance. We provide competitive baselines and open-source the code and the pre-trained models.

Luca Della Libera, Pooneh Mousavi, Salah Zaiem et al.

Modern multilingual automatic speech recognition (ASR) systems like Whisper have made it possible to transcribe audio in multiple languages with a single model. However, current state-of-the-art ASR models are typically evaluated on individual languages or in a multi-task setting, overlooking the challenge of continually learning new languages. There is insufficient research on how to add new languages without losing valuable information from previous data. Furthermore, existing continual learning benchmarks focus mostly on vision and language tasks, leaving continual learning for multilingual ASR largely unexplored. To bridge this gap, we propose CL-MASR, a benchmark designed for studying multilingual ASR in a continual learning setting. CL-MASR provides a diverse set of continual learning methods implemented on top of large-scale pretrained ASR models, along with common metrics to assess the effectiveness of learning new languages while addressing the issue of catastrophic forgetting. To the best of our knowledge, CL-MASR is the first continual learning benchmark for the multilingual ASR task. The code is available at https://github.com/speechbrain/benchmarks.

Zhepei Wang, Cem Subakan, Xilin Jiang et al. · uw

In this paper, we work on a sound recognition system that continually incorporates new sound classes. Our main goal is to develop a framework where the model can be updated without relying on labeled data. For this purpose, we propose adopting representation learning, where an encoder is trained using unlabeled data. This learning framework enables the study and implementation of a practically relevant use case where only a small amount of the labels is available in a continual learning context. We also make the empirical observation that a similarity-based representation learning method within this framework is robust to forgetting even if no explicit mechanism against forgetting is employed. We show that this approach obtains similar performance compared to several distillation-based continual learning methods when employed on self-supervised representation learning methods.

Luca Della Libera, Cem Subakan, Mirco Ravanelli

Neural audio codecs are at the core of modern conversational speech technologies, converting continuous speech into sequences of discrete tokens that can be processed by LLMs. However, existing codecs typically operate at fixed frame rates, allocating tokens uniformly in time and producing unnecessarily long sequences. In this work, we introduce DyCAST, a Dynamic Character-Aligned Speech Tokenizer that enables variable-frame-rate tokenization through soft character-level alignment and explicit duration modeling. DyCAST learns to associate tokens with character-level linguistic units during training and supports alignment-free inference with direct control over token durations at decoding time. To improve speech resynthesis quality at low frame rates, we further introduce a retrieval-augmented decoding mechanism that enhances reconstruction fidelity without increasing bitrate. Experiments show that DyCAST achieves competitive speech resynthesis quality and downstream performance while using significantly fewer tokens than fixed-frame-rate codecs. Code and checkpoints will be released publicly at https://github.com/lucadellalib/dycast.

Amirmohammad Sarfi, Zahra Karimpour, Muawiz Chaudhary et al. · mila

Recently, a number of iterative learning methods have been introduced to improve generalization. These typically rely on training for longer periods of time in exchange for improved generalization. LLF (later-layer-forgetting) is a state-of-the-art method in this category. It strengthens learning in early layers by periodically re-initializing the last few layers of the network. Our principal innovation in this work is to use Simulated annealing in EArly Layers (SEAL) of the network in place of re-initialization of later layers. Essentially, later layers go through the normal gradient descent process, while the early layers go through short stints of gradient ascent followed by gradient descent. Extensive experiments on the popular Tiny-ImageNet dataset benchmark and a series of transfer learning and few-shot learning tasks show that we outperform LLF by a significant margin. We further show that, compared to normal training, LLF features, although improving on the target task, degrade the transfer learning performance across all datasets we explored. In comparison, our method outperforms LLF across the same target datasets by a large margin. We also show that the prediction depth of our method is significantly lower than that of LLF and normal training, indicating on average better prediction performance.

Dominique Beaini, Shenyang Huang, Joao Alex Cunha et al.

Recently, pre-trained foundation models have enabled significant advancements in multiple fields. In molecular machine learning, however, where datasets are often hand-curated, and hence typically small, the lack of datasets with labeled features, and codebases to manage those datasets, has hindered the development of foundation models. In this work, we present seven novel datasets categorized by size into three distinct categories: ToyMix, LargeMix and UltraLarge. These datasets push the boundaries in both the scale and the diversity of supervised labels for molecular learning. They cover nearly 100 million molecules and over 3000 sparsely defined tasks, totaling more than 13 billion individual labels of both quantum and biological nature. In comparison, our datasets contain 300 times more data points than the widely used OGB-LSC PCQM4Mv2 dataset, and 13 times more than the quantum-only QM1B dataset. In addition, to support the development of foundational models based on our proposed datasets, we present the Graphium graph machine learning library which simplifies the process of building and training molecular machine learning models for multi-task and multi-level molecular datasets. Finally, we present a range of baseline results as a starting point of multi-task and multi-level training on these datasets. Empirically, we observe that performance on low-resource biological datasets show improvement by also training on large amounts of quantum data. This indicates that there may be potential in multi-task and multi-level training of a foundation model and fine-tuning it to resource-constrained downstream tasks.

Salah Zaiem, Youcef Kemiche, Titouan Parcollet et al.

Self-supervised learning (SSL) has recently allowed leveraging large datasets of unlabeled speech signals to reach impressive performance on speech tasks using only small amounts of annotated data. The high number of proposed approaches fostered the need and rise of extended benchmarks that evaluate their performance on a set of downstream tasks exploring various aspects of the speech signal. However, and while the number of considered tasks has been growing, most rely upon a single decoding architecture that maps the frozen SSL representations to the downstream labels. This work investigates the robustness of such benchmarking results to changes in the decoder architecture. Interestingly, it appears that varying the architecture of the downstream decoder leads to significant variations in the leaderboards of most tasks. Concerningly, our study reveals that benchmarking using limited decoders may cause a counterproductive increase in the sizes of the developed SSL models.

Francesco Paissan, Luca Della Libera, Zhepei Wang et al.

In this paper, we explore audio-editing with non-rigid text edits. We show that the proposed editing pipeline is able to create audio edits that remain faithful to the input audio. We explore text prompts that perform addition, style transfer, and in-painting. We quantitatively and qualitatively show that the edits are able to obtain results which outperform Audio-LDM, a recently released text-prompted audio generation model. Qualitative inspection of the results points out that the edits given by our approach remain more faithful to the input audio in terms of keeping the original onsets and offsets of the audio events.

Artem Ploujnikov, Mirco Ravanelli

End-to-end speech synthesis models directly convert the input characters into an audio representation (e.g., spectrograms). Despite their impressive performance, such models have difficulty disambiguating the pronunciations of identically spelled words. To mitigate this issue, a separate Grapheme-to-Phoneme (G2P) model can be employed to convert the characters into phonemes before synthesizing the audio. This paper proposes SoundChoice, a novel G2P architecture that processes entire sentences rather than operating at the word level. The proposed architecture takes advantage of a weighted homograph loss (that improves disambiguation), exploits curriculum learning (that gradually switches from word-level to sentence-level G2P), and integrates word embeddings from BERT (for further performance improvement). Moreover, the model inherits the best practices in speech recognition, including multi-task learning with Connectionist Temporal Classification (CTC) and beam search with an embedded language model. As a result, SoundChoice achieves a Phoneme Error Rate (PER) of 2.65% on whole-sentence transcription using data from LibriSpeech and Wikipedia. Index Terms grapheme-to-phoneme, speech synthesis, text-tospeech, phonetics, pronunciation, disambiguation.

Salah Zaiem, Youcef Kemiche, Titouan Parcollet et al.

Self-supervised learning (SSL) leverages large datasets of unlabeled speech to reach impressive performance with reduced amounts of annotated data. The high number of proposed approaches fostered the emergence of comprehensive benchmarks that evaluate their performance on a set of downstream tasks exploring various aspects of the speech signal. However, while the number of considered tasks has been growing, most proposals rely upon a single downstream architecture that maps the frozen SSL representations to the task labels. This study examines how benchmarking results are affected by changes in the probing head architecture. Interestingly, we found that altering the downstream architecture structure leads to significant fluctuations in the performance ranking of the evaluated models. Against common practices in speech SSL benchmarking, we evaluate larger-capacity probing heads, showing their impact on performance, inference costs, generalization and multi-level feature exploitation.

Salah Zaiem, Robin Algayres, Titouan Parcollet et al.

Self-supervised learning (SSL) has allowed substantial progress in Automatic Speech Recognition (ASR) performance in low-resource settings. In this context, it has been demonstrated that larger self-supervised feature extractors are crucial for achieving lower downstream ASR error rates. Thus, better performance might be sanctioned with longer inferences. This article explores different approaches that may be deployed during the fine-tuning to reduce the computations needed in the SSL encoder, leading to faster inferences. We adapt a number of existing techniques to common ASR settings and benchmark them, displaying performance drops and gains in inference times. Interestingly, we found that given enough downstream data, a simple downsampling of the input sequences outperforms the other methods with both low performance drops and high computational savings, reducing computations by 61.3% with an WER increase of only 0.81. Finally, we analyze the robustness of the comparison to changes in dataset conditions, revealing sensitivity to dataset size.

Sangeet Sagar, Mirco Ravanelli, Bernd Kiefer et al.

Despite the recent advancements in speech recognition, there are still difficulties in accurately transcribing conversational and emotional speech in noisy and reverberant acoustic environments. This poses a particular challenge in the search and rescue (SAR) domain, where transcribing conversations among rescue team members is crucial to support real-time decision-making. The scarcity of speech data and associated background noise in SAR scenarios make it difficult to deploy robust speech recognition systems. To address this issue, we have created and made publicly available a German speech dataset called RescueSpeech. This dataset includes real speech recordings from simulated rescue exercises. Additionally, we have released competitive training recipes and pre-trained models. Our study highlights that the performance attained by state-of-the-art methods in this challenging scenario is still far from reaching an acceptable level.

Ada Defne Tur, Adel Moumen, Mirco Ravanelli

Large Language Models (LLMs) have shown their ability to improve the performance of speech recognizers by effectively rescoring the n-best hypotheses generated during the beam search process. However, the best way to exploit recent generative instruction-tuned LLMs for hypothesis rescoring is still unclear. This paper proposes a novel method that uses instruction-tuned LLMs to dynamically expand the n-best speech recognition hypotheses with new hypotheses generated through appropriately-prompted LLMs. Specifically, we introduce a new zero-shot method for ASR n-best rescoring, which combines confidence scores, LLM sequence scoring, and prompt-based hypothesis generation. We compare Llama-3-Instruct, GPT-3.5 Turbo, and GPT-4 Turbo as prompt-based generators with Llama-3 as sequence scorer LLM. We evaluated our approach using different speech recognizers and observed significant relative improvement in the word error rate (WER) ranging from 5% to 25%.

Francesco Paissan, Cem Subakan, Mirco Ravanelli

In this paper, we introduce a new approach, called Posthoc Interpretation via Quantization (PIQ), for interpreting decisions made by trained classifiers. Our method utilizes vector quantization to transform the representations of a classifier into a discrete, class-specific latent space. The class-specific codebooks act as a bottleneck that forces the interpreter to focus on the parts of the input data deemed relevant by the classifier for making a prediction. Our model formulation also enables learning concepts by incorporating the supervision of pretrained annotation models such as state-of-the-art image segmentation models. We evaluated our method through quantitative and qualitative studies involving black-and-white images, color images, and audio. As a result of these studies we found that PIQ generates interpretations that are more easily understood by participants to our user studies when compared to several other interpretation methods in the literature.

Yingzhi Wang, Pooneh Mousavi, Artem Ploujnikov et al.

In audio and speech processing, tasks usually focus on either the audio or speech modality, even when both sounds and human speech are present in the same audio clip. Recent Auditory Large Language Models (ALLMs) have made it possible to process audio and speech simultaneously within a single model, leading to further considerations of joint audio-speech tasks. In this paper, we establish a novel benchmark to investigate how well ALLMs can perform joint audio-speech processing. Specifically, we introduce Joint Audio-Speech Co-Reasoning (JASCO), a novel task that unifies audio and speech processing, strictly requiring co-reasoning across both modalities. We also release a scene-reasoning dataset called "What Are They Doing". Additionally, we provide deeper insights into the models' behaviors by analyzing their dependence on each modality.

Giuseppe Alessio D'Inverno, Simone Brugiapaglia, Mirco Ravanelli

Graph Neural Networks (GNNs) have emerged as a powerful tool for data-driven learning on various graph domains. They are usually based on a message-passing mechanism and have gained increasing popularity for their intuitive formulation, which is closely linked to the Weisfeiler-Lehman (WL) test for graph isomorphism to which they have been proven equivalent in terms of expressive power. In this work, we establish new generalization properties and fundamental limits of GNNs in the context of learning so-called identity effects, i.e., the task of determining whether an object is composed of two identical components or not. Our study is motivated by the need to understand the capabilities of GNNs when performing simple cognitive tasks, with potential applications in computational linguistics and chemistry. We analyze two case studies: (i) two-letters words, for which we show that GNNs trained via stochastic gradient descent are unable to generalize to unseen letters when utilizing orthogonal encodings like one-hot representations; (ii) dicyclic graphs, i.e., graphs composed of two cycles, for which we present positive existence results leveraging the connection between GNNs and the WL test. Our theoretical analysis is supported by an extensive numerical study.

Jama Hussein Mohamud, Drew Wagner, Mirco Ravanelli

Mixture-of-Experts (MoE) layers increase model capacity by activating only a small subset of experts per token, and typically rely on a learned router to map hidden states to expert assignments. In this work, we ask whether a dedicated learned router is strictly necessary in the MoE settings we study. We propose Self-Routing, a parameter-free routing mechanism that uses a designated subspace of the token hidden state directly as expert logits, eliminating the router projection entirely while leaving the rest of the MoE layer unchanged. We evaluate Self-Routing on GPT-2-scale language modeling and ImageNet-1K classification by comparing it against a standard learned router, random-routing baselines, and dense non-MoE baselines. Our results show that Self-Routing remains competitive with the learned-router baseline while removing all dedicated routing parameters, and yields more balanced expert utilization, with about 17 % higher average normalized routing entropy and no explicit load-balancing loss. On ImageNet-1K with DeiT-S/16, Self-Routing also slightly improves over the corresponding learned-router MoE. These findings suggest that effective MoE routing can emerge from the hidden representation itself without requiring a separate learned router module.

Eleonora Mancini, Francesco Paissan, Mirco Ravanelli et al.

Neural networks are typically black-boxes that remain opaque with regards to their decision mechanisms. Several works in the literature have proposed post-hoc explanation methods to alleviate this issue. This paper proposes LMAC-TD, a post-hoc explanation method that trains a decoder to produce explanations directly in the time domain. This methodology builds upon the foundation of L-MAC, Listenable Maps for Audio Classifiers, a method that produces faithful and listenable explanations. We incorporate SepFormer, a popular transformer-based time-domain source separation architecture. We show through a user study that LMAC-TD significantly improves the audio quality of the produced explanations while not sacrificing from faithfulness.

Jihoon Jeong, Pooneh Mousavi, Mirco Ravanelli et al.

Large audio-language models (LALMs) can generate reasoning chains for their predictions, but it remains unclear whether these reasoning chains remain grounded in the input audio. In this paper, we propose an RL-based strategy that grounds the reasoning outputs of LALMs with explicit timestamp annotations referring to relevant segments of the audio signal. Our analysis shows that timestamp grounding leads the model to attend more strongly to audio tokens during reasoning generation. Experiments on four speech-based benchmark datasets demonstrate that our approach improves performance compared to both zero-shot reasoning and fine-tuning without timestamp grounding. Additionally, grounding amplifies desirable reasoning behaviors, such as region exploration, audiology verification, and consistency, underscoring the importance of grounding mechanisms for faithful multimodal reasoning.

Jingyi Li, Luca Della Libera, Mirco Ravanelli et al.

Many speech enhancement (SE) methods rely on continuous representations. Recently, discrete audio tokens have been explored to enable autoregressive generation for SE. However, it remains unclear whether discretization itself consistently improves SE performance. In this paper, we introduce LL-SDR, a token-based speech enhancement framework that explicitly leverages discretization to better separate speech and noise. Our first contribution is a Variance-Ordered Residual Vector Quantizer (VO-RVQ), designed to disentangle speech and noise distributions during tokenization. Second, we propose a latent-space discriminator to better align enhanced embeddings with semantic embeddings. Experiments show that LL-SDR outperforms continuous baselines and matches the performance of autoregressive token-based approaches, while enabling lightweight, low-latency speech enhancement in both reverberant and non-reverberant noisy environments. Demos and source code are available at our project websites.

Luca Della Libera, Jacopo Andreoli, Davide Dalle Pezze et al.

A crucial task in predictive maintenance is estimating the remaining useful life of physical systems. In the last decade, deep learning has improved considerably upon traditional model-based and statistical approaches in terms of predictive performance. However, in order to optimally plan maintenance operations, it is also important to quantify the uncertainty inherent to the predictions. This issue can be addressed by turning standard frequentist neural networks into Bayesian neural networks, which are naturally capable of providing confidence intervals around the estimates. Several methods exist for training those models. Researchers have focused mostly on parametric variational inference and sampling-based techniques, which notoriously suffer from limited approximation power and large computational burden, respectively. In this work, we use Stein variational gradient descent, a recently proposed algorithm for approximating intractable distributions that overcomes the drawbacks of the aforementioned techniques. In particular, we show through experimental studies on simulated run-to-failure turbofan engine degradation data that Bayesian deep learning models trained via Stein variational gradient descent consistently outperform with respect to convergence speed and predictive performance both the same models trained via parametric variational inference and their frequentist counterparts trained via backpropagation. Furthermore, we propose a method to enhance performance based on the uncertainty information provided by the Bayesian models. We release the source code at https://github.com/lucadellalib/bdl-rul-svgd.

Francesco Paissan, Luca Della Libera, Mirco Ravanelli et al.

Neural audio codecs provide compact discrete representations for speech generation and manipulation. However, most codecs organize tokens as frame-level sequences, making it difficult to study or intervene on global factors of variation. In this work, we propose the Latent Audio Tokenizer for Token-space Editing (LATTE) that appends a fixed set of learnable latent tokens to the audio feature sequence and retains only these tokens for quantization and decoding. This design produces a compact, non-temporally aligned bottleneck in which each token can aggregate global information across the full utterance. We show that the resulting tokenizer preserves competitive reconstruction quality in low-bitrate speech coding settings while enabling simple token-space interventions. In particular, we find that swapping selected latent token positions between utterances can modify global attributes, such as speaker identity and background noise, and we evaluate these interventions on voice conversion and denoising tasks. Our results suggest that compact latent audio tokenizers can support controllable audio manipulation without supervision in task-specific editing models.

Luca Della Libera, Cem Subakan, Mirco Ravanelli

Neural audio codecs are a fundamental component of modern generative audio pipelines. Although recent codecs achieve strong low-bitrate reconstruction and provide powerful representations for downstream tasks, most are non-streamable, limiting their use in real-time applications. We present FocalCodec-Stream, a hybrid codec based on focal modulation that compresses speech into a single binary codebook at 0.55 - 0.80 kbps with a theoretical latency of 80 ms. Our approach combines multi-stage causal distillation of WavLM with targeted architectural improvements, including a lightweight refiner module that enhances quality under latency constraints. Experiments show that FocalCodec-Stream outperforms existing streamable codecs at comparable bitrates, while preserving both semantic and acoustic information. The result is a favorable trade-off between reconstruction quality, downstream task performance, latency, and efficiency. Code and checkpoints will be released at https://github.com/lucadellalib/focalcodec.

Mirco Ravanelli, Titouan Parcollet, Adel Moumen et al.

SpeechBrain is an open-source Conversational AI toolkit based on PyTorch, focused particularly on speech processing tasks such as speech recognition, speech enhancement, speaker recognition, text-to-speech, and much more. It promotes transparency and replicability by releasing both the pre-trained models and the complete "recipes" of code and algorithms required for training them. This paper presents SpeechBrain 1.0, a significant milestone in the evolution of the toolkit, which now has over 200 recipes for speech, audio, and language processing tasks, and more than 100 models available on Hugging Face. SpeechBrain 1.0 introduces new technologies to support diverse learning modalities, Large Language Model (LLM) integration, and advanced decoding strategies, along with novel models, tasks, and modalities. It also includes a new benchmark repository, offering researchers a unified platform for evaluating models across diverse tasks.

Mirco Ravanelli, Titouan Parcollet, Peter Plantinga et al.

SpeechBrain is an open-source and all-in-one speech toolkit. It is designed to facilitate the research and development of neural speech processing technologies by being simple, flexible, user-friendly, and well-documented. This paper describes the core architecture designed to support several tasks of common interest, allowing users to naturally conceive, compare and share novel speech processing pipelines. SpeechBrain achieves competitive or state-of-the-art performance in a wide range of speech benchmarks. It also provides training recipes, pretrained models, and inference scripts for popular speech datasets, as well as tutorials which allow anyone with basic Python proficiency to familiarize themselves with speech technologies.

Loren Lugosch, Piyush Papreja, Mirco Ravanelli et al.

This paper introduces Timers and Such, a new open source dataset of spoken English commands for common voice control use cases involving numbers. We describe the gap in existing spoken language understanding datasets that Timers and Such fills, the design and creation of the dataset, and experiments with a number of ASR-based and end-to-end baseline models, the code for which has been made available as part of the SpeechBrain toolkit.

Mirco Ravanelli, Titouan Parcollet, Yoshua Bengio

The availability of open-source software is playing a remarkable role in the popularization of speech recognition and deep learning. Kaldi, for instance, is nowadays an established framework used to develop state-of-the-art speech recognizers. PyTorch is used to build neural networks with the Python language and has recently spawn tremendous interest within the machine learning community thanks to its simplicity and flexibility. The PyTorch-Kaldi project aims to bridge the gap between these popular toolkits, trying to inherit the efficiency of Kaldi and the flexibility of PyTorch. PyTorch-Kaldi is not only a simple interface between these software, but it embeds several useful features for developing modern speech recognizers. For instance, the code is specifically designed to naturally plug-in user-defined acoustic models. As an alternative, users can exploit several pre-implemented neural networks that can be customized using intuitive configuration files. PyTorch-Kaldi supports multiple feature and label streams as well as combinations of neural networks, enabling the use of complex neural architectures. The toolkit is publicly-released along with a rich documentation and is designed to properly work locally or on HPC clusters. Experiments, that are conducted on several datasets and tasks, show that PyTorch-Kaldi can effectively be used to develop modern state-of-the-art speech recognizers.

Pooneh Mousavi, Gallil Maimon, Adel Moumen et al.

Discrete audio tokens are compact representations that aim to preserve perceptual quality, phonetic content, and speaker characteristics while enabling efficient storage and inference, as well as competitive performance across diverse downstream tasks. They provide a practical alternative to continuous features, enabling the integration of speech and audio into modern large language models (LLMs). As interest in token-based audio processing grows, various tokenization methods have emerged, and several surveys have reviewed the latest progress in the field. However, existing studies often focus on specific domains or tasks and lack a unified comparison across various benchmarks. This paper presents a systematic review and benchmark of discrete audio tokenizers, covering three domains: speech, music, and general audio. We propose a taxonomy of tokenization approaches based on encoder-decoder, quantization techniques, training paradigm, streamability, and application domains. We evaluate tokenizers on multiple benchmarks for reconstruction, downstream performance, and acoustic language modeling, and analyze trade-offs through controlled ablation studies. Our findings highlight key limitations, practical considerations, and open challenges, providing insight and guidance for future research in this rapidly evolving area. For more information, including our main results and tokenizer database, please refer to our website: https://poonehmousavi.github.io/dates-website/.

Seyed Mahed Mousavi, Gabriel Roccabruna, Simone Alghisi et al.

Large Pre-Trained Language Models have demonstrated state-of-the-art performance in different downstream tasks, including dialogue state tracking and end-to-end response generation. Nevertheless, most of the publicly available datasets and benchmarks on task-oriented dialogues focus on written conversations. Consequently, the robustness of the developed models to spoken interactions is unknown. In this work, we have evaluated the performance of LLMs for spoken task-oriented dialogues on the DSTC11 test sets. Due to the lack of proper spoken dialogue datasets, we have automatically transcribed a development set of spoken dialogues with a state-of-the-art ASR engine. We have characterized the ASR-error types and their distributions and simulated these errors in a large dataset of dialogues. We report the intrinsic (perplexity) and extrinsic (human evaluation) performance of fine-tuned GPT-2 and T5 models in two subtasks of response generation and dialogue state tracking, respectively. The results show that LLMs are not robust to spoken noise by default, however, fine-tuning/training such models on a proper dataset of spoken TODs can result in a more robust performance.

Luca Della Libera, Francesco Paissan, Cem Subakan et al.

Large language models have revolutionized natural language processing through self-supervised pretraining on massive datasets. Inspired by this success, researchers have explored adapting these methods to speech by discretizing continuous audio into tokens using neural audio codecs. However, existing approaches face limitations, including high bitrates, the loss of either semantic or acoustic information, and the reliance on multi-codebook designs when trying to capture both, which increases architectural complexity for downstream tasks. To address these challenges, we introduce FocalCodec, an efficient low-bitrate codec based on focal modulation that utilizes a single binary codebook to compress speech between 0.16 and 0.65 kbps. FocalCodec delivers competitive performance in speech resynthesis and voice conversion at lower bitrates than the current state-of-the-art, while effectively handling multilingual speech and noisy environments. Evaluation on downstream tasks shows that FocalCodec successfully preserves sufficient semantic and acoustic information, while also being well-suited for generative modeling. Demo samples and code are available at https://lucadellalib.github.io/focalcodec-web/.

Eleonora Mancini, Francesco Paissan, Paolo Torroni et al.

Speech impairments in Parkinson's disease (PD) provide significant early indicators for diagnosis. While models for speech-based PD detection have shown strong performance, their interpretability remains underexplored. This study systematically evaluates several explainability methods to identify PD-specific speech features, aiming to support the development of accurate, interpretable models for clinical decision-making in PD diagnosis and monitoring. Our methodology involves (i) obtaining attributions and saliency maps using mainstream interpretability techniques, (ii) quantitatively evaluating the faithfulness of these maps and their combinations obtained via union and intersection through a range of established metrics, and (iii) assessing the information conveyed by the saliency maps for PD detection from an auxiliary classifier. Our results reveal that, while explanations are aligned with the classifier, they often fail to provide valuable information for domain experts.

Luca Zampierin, Ghouthi Boukli Hacene, Bac Nguyen et al.

Self-supervised learning (SSL) has achieved remarkable success across various speech-processing tasks. To enhance its efficiency, previous works often leverage the use of compression techniques. A notable recent attempt is DPHuBERT, which applies joint knowledge distillation (KD) and structured pruning to learn a significantly smaller SSL model. In this paper, we contribute to this research domain by introducing SKILL, a novel method that conducts distillation across groups of layers instead of distilling individual arbitrarily selected layers within the teacher network. The identification of the layers to distill is achieved through a hierarchical clustering procedure applied to layer similarity measures. Extensive experiments demonstrate that our distilled version of WavLM Base+ not only outperforms DPHuBERT but also achieves state-of-the-art results in the 30M parameters model class across several SUPERB tasks.

Luca Della Libera, Cem Subakan, Mirco Ravanelli

The increasing success of deep neural networks has raised concerns about their inherent black-box nature, posing challenges related to interpretability and trust. While there has been extensive exploration of interpretation techniques in vision and language, interpretability in the audio domain has received limited attention, primarily focusing on post-hoc explanations. This paper addresses the problem of interpretability by-design in the audio domain by utilizing the recently proposed attention-free focal modulation networks (FocalNets). We apply FocalNets to the task of environmental sound classification for the first time and evaluate their interpretability properties on the popular ESC-50 dataset. Our method outperforms a similarly sized vision transformer both in terms of accuracy and interpretability. Furthermore, it is competitive against PIQ, a method specifically designed for post-hoc interpretation in the audio domain.

Luca Della Libera, Cem Subakan, Mirco Ravanelli

In speech processing pipelines, improving the quality and intelligibility of real-world recordings is crucial. While supervised regression is the primary method for speech enhancement, audio tokenization is emerging as a promising alternative for a smooth integration with other modalities. However, research on speech enhancement using discrete representations is still limited. Previous work has mainly focused on semantic tokens, which tend to discard key acoustic details such as speaker identity. Additionally, these studies typically employ non-autoregressive models, assuming conditional independence of outputs and overlooking the potential improvements offered by autoregressive modeling. To address these gaps we: 1) conduct a comprehensive study of the performance of acoustic tokens for speech enhancement, including the effect of bitrate and noise strength; 2) introduce a novel transducer-based autoregressive architecture specifically designed for this task. Experiments on VoiceBank and Libri1Mix datasets show that acoustic tokens outperform semantic tokens in terms of preserving speaker identity, and that our autoregressive approach can further improve performance. Nevertheless, we observe that discrete representations still fall short compared to continuous ones, highlighting the need for further research in this area.

Pooneh Mousavi, Shubham Gupta, Cem Subakan et al.

Foundation models based on large language models (LLMs) have shown great success in handling various tasks and modalities. However, adapting these models for general-purpose audio-language tasks is challenging due to differences in acoustic environments and task variations. In this work, we introduce LiSTEN Learning Soft Token Embeddings for Neural Audio LLMs), a framework for adapting LLMs to speech and audio tasks. LiSTEN uses a dynamic prompt selection strategy with learnable key-value pairs, allowing the model to balance general and task-specific knowledge while avoiding overfitting in a multitask setting. Our approach reduces dependence on large-scale ASR or captioning datasets, achieves competitive performance with fewer trainable parameters, and simplifies training by using a single-stage process. Additionally, LiSTEN enhances interpretability by analyzing the diversity and overlap of selected prompts across different tasks.

Maab Elrashid, Anthony Deschênes, Cem Subakan et al.

Interpretability is essential for user trust in real-world anomaly detection applications. However, deep learning models, despite their strong performance, often lack transparency. In this work, we study the interpretability of autoencoder-based models for audio anomaly detection, by comparing a standard autoencoder (AE) with a mask autoencoder (MAE) in terms of detection performance and interpretability. We applied several attribution methods, including error maps, saliency maps, SmoothGrad, Integrated Gradients, GradSHAP, and Grad-CAM. Although MAE shows a slightly lower detection, it consistently provides more faithful and temporally precise explanations, suggesting a better alignment with true anomalies. To assess the relevance of the regions highlighted by the explanation method, we propose a perturbation-based faithfulness metric that replaces them with their reconstructions to simulate normal input. Our findings, based on experiments in a real industrial scenario, highlight the importance of incorporating interpretability into anomaly detection pipelines and show that masked training improves explanation quality without compromising performance.

Luca Della Libera, Cem Subakan, Mirco Ravanelli

Large language models show that simple autoregressive training can yield scalable and coherent generation, but extending this paradigm to speech remains challenging due to the entanglement of semantic and acoustic information. Most existing speech language models rely on text supervision, hierarchical token streams, or complex hybrid architectures, departing from the single-stream generative pretraining paradigm that has proven effective in text. In this work, we introduce WavSLM, a speech language model trained by quantizing and distilling self-supervised WavLM representations into a single codebook and optimizing an autoregressive next-chunk prediction objective. WavSLM jointly models semantic and acoustic information within a single token stream without text supervision or text pretraining. Despite its simplicity, it achieves competitive performance on consistency benchmarks and speech generation while using fewer parameters, less training data, and supporting streaming inference. Demo samples are available at https://lucadellalib.github.io/wavslm-web/.

Lovenya Jain, Pooneh Mousavi, Mirco Ravanelli et al.

Faithfulness measures whether chain-of-thought (CoT) representations accurately reflect a model's decision process and can be used as reliable explanations. Prior work has shown that CoTs from text-based LLMs are often unfaithful. This question has not been explored for large audio-language models (LALMs), where faithfulness is critical for safety-sensitive applications. Reasoning in LALMs is also more challenging, as models must first extract relevant clues from audio before reasoning over them. In this paper, we investigate the faithfulness of CoTs produced by several LALMs by applying targeted interventions, including paraphrasing, filler token injection, early answering, and introducing mistakes, on two challenging reasoning datasets: SAKURA and MMAR. After going through the aforementioned interventions across several datasets and tasks, our experiments suggest that, LALMs generally produce CoTs that appear to be faithful to their underlying decision processes.

Matthieu Cervera, Francesco Paissan, Mirco Ravanelli et al.

Free-form, text-based audio editing remains a persistent challenge, despite progress in inversion-based neural methods. Current approaches rely on slow inversion procedures, limiting their practicality. We present a virtual-consistency based audio editing system that bypasses inversion by adapting the sampling process of diffusion models. Our pipeline is model-agnostic, requiring no fine-tuning or architectural changes, and achieves substantial speed-ups over recent neural editing baselines. Crucially, it achieves this efficiency without compromising quality, as demonstrated by quantitative benchmarks and a user study involving 16 participants.

Peter Plantinga, Jen-Kai Chen, Roozbeh Sattari et al. · mila

Speech holds promise as a cost-effective and non-invasive biomarker for neurological conditions such as Parkinson's disease (PD). While deep learning systems trained on raw audio can find subtle signals not available from hand-crafted features, their black-box nature hinders clinical adoption. To address this, we apply sparse autoencoders (SAEs) to uncover interpretable internal representations from a speech-based PD detection system. We introduce a novel mask-based activation for adapting SAEs to small biomedical datasets, creating sparse disentangled dictionary representations. These dictionary entries are found to have strong associations with characteristic articulatory deficits in PD speech, such as reduced spectral flux and increased spectral flatness in the low-energy regions highlighted by the model attention. We further show that the spectral flux is related to volumetric measurements of the putamen from MRI scans, demonstrating the potential of SAEs to reveal clinically relevant biomarkers for disease monitoring and diagnosis.

Peter Plantinga, Briac Cordelle, Dominique Louër et al.

Using speech samples as a biomarker is a promising avenue for detecting and monitoring the progression of Parkinson's disease (PD), but there is considerable disagreement in the literature about how best to collect and analyze such data. Early research in detecting PD from speech used a sustained vowel phonation (SVP) task, while some recent research has explored recordings of more cognitively demanding tasks. To assess the role of language in PD detection, we tested pretrained models with varying data types and pretraining objectives and found that (1) text-only models match the performance of vocal-feature models, (2) multilingual Whisper outperforms self-supervised models whereas monolingual Whisper does worse, and (3) AudioSet pretraining improves performance on SVP but not spontaneous speech. These findings together highlight the critical role of language for the early detection of Parkinson's disease.

Yingzhi Wang, Anas Alhmoud, Saad Alsahly et al.

OpenAI's Whisper has achieved significant success in Automatic Speech Recognition. However, it has consistently been found to exhibit hallucination issues, particularly in non-speech segments, which limits its broader application in complex industrial settings. In this paper, we introduce a novel method to reduce Whisper's hallucination on non-speech segments without using any pre- or post-possessing techniques. Specifically, we benchmark the contribution of each self-attentional head in the Whisper-large-v3 decoder to the hallucination problem by performing a head-wise mask. Our findings reveal that only 3 of the 20 heads account for over 75% of the hallucinations on the UrbanSound dataset. We then fine-tune these three crazy heads using a collection of non-speech data. The results show that our best fine-tuned model, namely Calm-Whisper, achieves over 80% reduction in non-speech hallucination with only less than 0.1% WER degradation on LibriSpeech test-clean and test-other.

Pooneh Mousavi, Luca Della Libera, Jarod Duret et al.

Discrete audio tokens have recently gained considerable attention for their potential to connect audio and language processing, enabling the creation of modern multimodal large language models. Ideal audio tokens must effectively preserve phonetic and semantic content along with paralinguistic information, speaker identity, and other details. While several types of audio tokens have been recently proposed, identifying the optimal tokenizer for various tasks is challenging due to the inconsistent evaluation settings in existing studies. To address this gap, we release the Discrete Audio and Speech Benchmark (DASB), a comprehensive leaderboard for benchmarking discrete audio tokens across a wide range of discriminative tasks, including speech recognition, speaker identification and verification, emotion recognition, keyword spotting, and intent classification, as well as generative tasks such as speech enhancement, separation, and text-to-speech. Our results show that, on average, semantic tokens outperform compression tokens across most discriminative and generative tasks. However, the performance gap between semantic tokens and standard continuous representations remains substantial, highlighting the need for further research in this field.

Pooneh Mousavi, Jarod Duret, Salah Zaiem et al.

Discrete audio tokens have recently gained attention for their potential to bridge the gap between audio and language processing. Ideal audio tokens must preserve content, paralinguistic elements, speaker identity, and many other audio details. Current audio tokenization methods fall into two categories: Semantic tokens, acquired through quantization of Self-Supervised Learning (SSL) models, and Neural compression-based tokens (codecs). Although previous studies have benchmarked codec models to identify optimal configurations, the ideal setup for quantizing pretrained SSL models remains unclear. This paper explores the optimal configuration of semantic tokens across discriminative and generative tasks. We propose a scalable solution to train a universal vocoder across multiple SSL layers. Furthermore, an attention mechanism is employed to identify task-specific influential layers, enhancing the adaptability and performance of semantic tokens in diverse audio applications.

Francesco Paissan, Mirco Ravanelli, Cem Subakan

Despite the impressive performance of deep learning models across diverse tasks, their complexity poses challenges for interpretation. This challenge is particularly evident for audio signals, where conveying interpretations becomes inherently difficult. To address this issue, we introduce Listenable Maps for Audio Classifiers (L-MAC), a posthoc interpretation method that generates faithful and listenable interpretations. L-MAC utilizes a decoder on top of a pretrained classifier to generate binary masks that highlight relevant portions of the input audio. We train the decoder with a loss function that maximizes the confidence of the classifier decision on the masked-in portion of the audio while minimizing the probability of model output for the masked-out portion. Quantitative evaluations on both in-domain and out-of-domain data demonstrate that L-MAC consistently produces more faithful interpretations than several gradient and masking-based methodologies. Furthermore, a user study confirms that, on average, users prefer the interpretations generated by the proposed technique.

Cem Subakan, Mirco Ravanelli, Samuele Cornell et al.

Transformers have enabled impressive improvements in deep learning. They often outperform recurrent and convolutional models in many tasks while taking advantage of parallel processing. Recently, we proposed the SepFormer, which obtains state-of-the-art performance in speech separation with the WSJ0-2/3 Mix datasets. This paper studies in-depth Transformers for speech separation. In particular, we extend our previous findings on the SepFormer by providing results on more challenging noisy and noisy-reverberant datasets, such as LibriMix, WHAM!, and WHAMR!. Moreover, we extend our model to perform speech enhancement and provide experimental evidence on denoising and dereverberation tasks. Finally, we investigate, for the first time in speech separation, the use of efficient self-attention mechanisms such as Linformers, Lonformers, and ReFormers. We found that they reduce memory requirements significantly. For example, we show that the Reformer-based attention outperforms the popular Conv-TasNet model on the WSJ0-2Mix dataset while being faster at inference and comparable in terms of memory consumption.

Cheng Yu, Szu-Wei Fu, Tsun-An Hsieh et al.

Although deep learning (DL) has achieved notable progress in speech enhancement (SE), further research is still required for a DL-based SE system to adapt effectively and efficiently to particular speakers. In this study, we propose a novel meta-learning-based speaker-adaptive SE approach (called OSSEM) that aims to achieve SE model adaptation in a one-shot manner. OSSEM consists of a modified transformer SE network and a speaker-specific masking (SSM) network. In practice, the SSM network takes an enrolled speaker embedding extracted using ECAPA-TDNN to adjust the input noisy feature through masking. To evaluate OSSEM, we designed a modified Voice Bank-DEMAND dataset, in which one utterance from the testing set was used for model adaptation, and the remaining utterances were used for testing the performance. Moreover, we set restrictions allowing the enhancement process to be conducted in real time, and thus designed OSSEM to be a causal SE system. Experimental results first show that OSSEM can effectively adapt a pretrained SE model to a particular speaker with only one utterance, thus yielding improved SE results. Meanwhile, OSSEM exhibits a competitive performance compared to state-of-the-art causal SE systems.

Cem Subakan, Mirco Ravanelli, Samuele Cornell et al.

In recent years, deep learning based source separation has achieved impressive results. Most studies, however, still evaluate separation models on synthetic datasets, while the performance of state-of-the-art techniques on in-the-wild speech data remains an open question. This paper contributes to fill this gap in two ways. First, we release the REAL-M dataset, a crowd-sourced corpus of real-life mixtures. Secondly, we address the problem of performance evaluation of real-life mixtures, where the ground truth is not available. We bypass this issue by carefully designing a blind Scale-Invariant Signal-to-Noise Ratio (SI-SNR) neural estimator. Through a user study, we show that our estimator reliably evaluates the separation performance on real mixtures. The performance predictions of the SI-SNR estimator indeed correlate well with human opinions. Moreover, we observe that the performance trends predicted by our estimator on the REAL-M dataset closely follow those achieved on synthetic benchmarks when evaluating popular speech separation models.