Alessio Brutti

Giovanni Morrone, Samuele Cornell, Luca Serafini et al. · cmu

Recent works show that speech separation guided diarization (SSGD) is an increasingly promising direction, mainly thanks to the recent progress in speech separation. It performs diarization by first separating the speakers and then applying voice activity detection (VAD) on each separated stream. In this work we conduct an in-depth study of SSGD in the conversational telephone speech (CTS) domain, focusing mainly on low-latency streaming diarization applications. We consider three state-of-the-art speech separation (SSep) algorithms and study their performance both in online and offline scenarios, considering non-causal and causal implementations as well as continuous SSep (CSS) windowed inference. We compare different SSGD algorithms on two widely used CTS datasets: CALLHOME and Fisher Corpus (Part 1 and 2) and evaluate both separation and diarization performance. To improve performance, a novel, causal and computationally efficient leakage removal algorithm is proposed, which significantly decreases false alarms. We also explore, for the first time, fully end-to-end SSGD integration between SSep and VAD modules. Crucially, this enables fine-tuning on real-world data for which oracle speakers sources are not available. In particular, our best model achieves 8.8% DER on CALLHOME, which outperforms the current state-of-the-art end-to-end neural diarization model, despite being trained on an order of magnitude less data and having significantly lower latency, i.e., 0.1 vs. 1 s. Finally, we also show that the separated signals can be readily used also for automatic speech recognition, reaching performance close to using oracle sources in some configurations.

Umberto Cappellazzo, Enrico Fini, Muqiao Yang et al. · cmu

Recently, neural networks have shown impressive progress across diverse fields, with speech processing being no exception. However, recent breakthroughs in this area require extensive offline training using large datasets and tremendous computing resources. Unfortunately, these models struggle to retain their previously acquired knowledge when learning new tasks continually, and retraining from scratch is almost always impractical. In this paper, we investigate the problem of learning sequence-to-sequence models for spoken language understanding in a class-incremental learning (CIL) setting and we propose COCONUT, a CIL method that relies on the combination of experience replay and contrastive learning. Through a modified version of the standard supervised contrastive loss applied only to the rehearsal samples, COCONUT preserves the learned representations by pulling closer samples from the same class and pushing away the others. Moreover, we leverage a multimodal contrastive loss that helps the model learn more discriminative representations of the new data by aligning audio and text features. We also investigate different contrastive designs to combine the strengths of the contrastive loss with teacher-student architectures used for distillation. Experiments on two established SLU datasets reveal the effectiveness of our proposed approach and significant improvements over the baselines. We also show that COCONUT can be combined with methods that operate on the decoder side of the model, resulting in further metrics improvements.

Umberto Cappellazzo, Minsu Kim, Honglie Chen et al.

Multimodal large language models (MLLMs) have recently become a focal point of research due to their formidable multimodal understanding capabilities. For example, in the audio and speech domains, an LLM can be equipped with (automatic) speech recognition (ASR) abilities by just concatenating the audio tokens, computed with an audio encoder, and the text tokens to achieve state-of-the-art results. On the contrary, tasks like visual and audio-visual speech recognition (VSR/AVSR), which also exploit noise-invariant lip movement information, have received little or no attention. To bridge this gap, we propose Llama-AVSR, a new MLLM with strong audio-visual speech recognition capabilities. It leverages pre-trained audio and video encoders to produce modality-specific tokens which, together with the text tokens, are processed by a pre-trained LLM (e.g., Llama3.1-8B) to yield the resulting response in an auto-regressive fashion. Llama-AVSR requires a small number of trainable parameters as only modality-specific projectors and LoRA modules are trained whereas the multi-modal encoders and LLM are kept frozen. We evaluate our proposed approach on LRS3, the largest public AVSR benchmark, and we achieve new state-of-the-art results for the tasks of ASR and AVSR with a WER of 0.79% and 0.77%, respectively. To bolster our results, we investigate the key factors that underpin the effectiveness of Llama-AVSR: the choice of the pre-trained encoders and LLM, the efficient integration of LoRA modules, and the optimal performance-efficiency trade-off obtained via modality-aware compression rates.

Umberto Cappellazzo, Daniele Falavigna, Alessio Brutti

Continual learning refers to a dynamical framework in which a model receives a stream of non-stationary data over time and must adapt to new data while preserving previously acquired knowledge. Unluckily, neural networks fail to meet these two desiderata, incurring the so-called catastrophic forgetting phenomenon. Whereas a vast array of strategies have been proposed to attenuate forgetting in the computer vision domain, for speech-related tasks, on the other hand, there is a dearth of works. In this paper, we consider the joint use of rehearsal and knowledge distillation (KD) approaches for spoken language understanding under a class-incremental learning scenario. We report on multiple KD combinations at different levels in the network, showing that combining feature-level and predictions-level KDs leads to the best results. Finally, we provide an ablation study on the effect of the size of the rehearsal memory that corroborates the efficacy of our approach for low-resource devices.

Mohamed Nabih Ali, Francesco Paissan, Daniele Falavigna et al.

Recently, several very effective neural approaches for single-channel speech separation have been presented in the literature. However, due to the size and complexity of these models, their use on low-resource devices, e.g. for hearing aids, and earphones, is still a challenge and established solutions are not available yet. Although approaches based on either pruning or compressing neural models have been proposed, the design of a model architecture suitable for a certain application domain often requires heuristic procedures not easily portable to different low-resource platforms. Given the modular nature of the well-known Conv-Tasnet speech separation architecture, in this paper we consider three parameters that directly control the overall size of the model, namely: the number of residual blocks, the number of repetitions of the separation blocks and the number of channels in the depth-wise convolutions, and experimentally evaluate how they affect the speech separation performance. In particular, experiments carried out on the Libri2Mix show that the number of dilated 1D-Conv blocks is the most critical parameter and that the usage of extra-dilation in the residual blocks allows reducing the performance drop.

Mohamed Nabih Ali, Alessio Brutti, Daniele Falavigna

Intent classification is a fundamental task in the spoken language understanding field that has recently gained the attention of the scientific community, mainly because of the feasibility of approaching it with end-to-end neural models. In this way, avoiding using intermediate steps, i.e. automatic speech recognition, is possible, thus the propagation of errors due to background noise, spontaneous speech, speaking styles of users, etc. Towards the development of solutions applicable in real scenarios, it is interesting to investigate how environmental noise and related noise reduction techniques to address the intent classification task with end-to-end neural models. In this paper, we experiment with a noisy version of the fluent speech command data set, combining the intent classifier with a time-domain speech enhancement solution based on Wave-U-Net and considering different training strategies. Experimental results reveal that, for this task, the use of speech enhancement greatly improves the classification accuracy in noisy conditions, in particular when the classification model is trained on enhanced signals.

Nithin Rao Koluguri, Monica Sekoyan, George Zelenfroynd et al. · nvidia

Multi-task and multilingual approaches benefit large models, yet speech processing for low-resource languages remains underexplored due to data scarcity. To address this, we present Granary, a large-scale collection of speech datasets for recognition and translation across 25 European languages. This is the first open-source effort at this scale for both transcription and translation. We enhance data quality using a pseudo-labeling pipeline with segmentation, two-pass inference, hallucination filtering, and punctuation restoration. We further generate translation pairs from pseudo-labeled transcriptions using EuroLLM, followed by a data filtration pipeline. Designed for efficiency, our pipeline processes vast amount of data within hours. We assess models trained on processed data by comparing their performance on previously curated datasets for both high- and low-resource languages. Our findings show that these models achieve similar performance using approx. 50% less data. Dataset will be made available at https://hf.co/datasets/nvidia/Granary

Sara Papi, Marco Gaido, Luisa Bentivogli et al.

The development of speech foundation models (SFMs) like Whisper and SeamlessM4T has significantly advanced the field of speech processing. However, their closed nature--with inaccessible training data and code--poses major reproducibility and fair evaluation challenges. While other domains have made substantial progress toward open science by developing fully transparent models trained on open-source (OS) code and data, similar efforts in speech remain limited. To fill this gap, we introduce FAMA, the first family of open science SFMs for English and Italian, trained on 150k+ hours of OS speech data. Moreover, we present a new dataset containing 16k hours of cleaned and pseudo-labeled speech for both languages. Results show that FAMA achieves competitive performance compared to existing SFMs while being up to 8 times faster. All artifacts, including code, datasets, and models, are released under OS-compliant licenses, promoting openness in speech technology research.

Vandana Rajan, Alessio Brutti, Andrea Cavallaro

Humans express their emotions via facial expressions, voice intonation and word choices. To infer the nature of the underlying emotion, recognition models may use a single modality, such as vision, audio, and text, or a combination of modalities. Generally, models that fuse complementary information from multiple modalities outperform their uni-modal counterparts. However, a successful model that fuses modalities requires components that can effectively aggregate task-relevant information from each modality. As cross-modal attention is seen as an effective mechanism for multi-modal fusion, in this paper we quantify the gain that such a mechanism brings compared to the corresponding self-attention mechanism. To this end, we implement and compare a cross-attention and a self-attention model. In addition to attention, each model uses convolutional layers for local feature extraction and recurrent layers for global sequential modelling. We compare the models using different modality combinations for a 7-class emotion classification task using the IEMOCAP dataset. Experimental results indicate that albeit both models improve upon the state-of-the-art in terms of weighted and unweighted accuracy for tri- and bi-modal configurations, their performance is generally statistically comparable. The code to replicate the experiments is available at https://github.com/smartcameras/SelfCrossAttn

Umberto Cappellazzo, Daniele Falavigna, Alessio Brutti

Mixture of Experts (MoE) architectures have recently started burgeoning due to their ability to scale model's capacity while maintaining the computational cost affordable. Furthermore, they can be applied to both Transformers and State Space Models, the current state-of-the-art models in numerous fields. While MoE has been mostly investigated for the pre-training stage, its use in parameter-efficient transfer learning settings is under-explored. To narrow this gap, this paper attempts to demystify the use of MoE for parameter-efficient fine-tuning of Audio Spectrogram Transformers to audio and speech downstream tasks. Specifically, we propose Soft Mixture of Adapters (Soft-MoA). It exploits adapters as the experts and, leveraging the recent Soft MoE method, it relies on a soft assignment between the input tokens and experts to keep the computational time limited. Extensive experiments across 4 benchmarks demonstrate that Soft-MoA outperforms the single adapter method and performs on par with the dense MoA counterpart. We finally present ablation studies on key elements of Soft-MoA, showing for example that Soft-MoA achieves better scaling with more experts, as well as ensuring that all experts contribute to the computation of the output tokens, thus dispensing with the expert imbalance issue.

Umberto Cappellazzo, Minsu Kim, Stavros Petridis et al.

Audio-Visual Speech Recognition (AVSR) enhances robustness in noisy environments by integrating visual cues. While recent advances integrate Large Language Models (LLMs) into AVSR, their high computational cost hinders deployment in resource-constrained settings. To address this, we propose Llama-SMoP, an efficient Multimodal LLM that employs a Sparse Mixture of Projectors (SMoP) module to scale model capacity without increasing inference costs. By incorporating sparsely-gated mixture-of-experts (MoE) projectors, Llama-SMoP enables the use of smaller LLMs while maintaining strong performance. We explore three SMoP configurations and show that Llama-SMoP DEDR (Disjoint-Experts, Disjoint-Routers), which uses modality-specific routers and experts, achieves superior performance on ASR, VSR, and AVSR tasks. Ablation studies confirm its effectiveness in expert activation, scalability, and noise robustness.

Lorenzo Concina, Jordi Luque, Alessio Brutti et al.

In this paper, we present our studies and experiments carried out for the task 1 of the Challenge and Workshop on Multilingual Conversational Speech Language Model (MLC-SLM), which focuses on advancing multilingual conversational speech recognition through the development of speech language models architectures. Given the increasing relevance of real-world conversational data for building robust Spoken Dialogue Systems, we explore three approaches to multilingual ASR. First, we conduct an evaluation of the official baseline to better understand its strengths and limitations, by training two projectors (linear and qformer) with different foundation models. Second we leverage the SLAM-ASR framework to train a custom multilingual linear projector. Finally we investigate the role of contrastive learning and the extended conversational context in enhancing the robustness of recognition.

Mohamed Nabih Ali, Daniele Falavigna, Alessio Brutti

Multilingual automatic speech recognition (ASR) remains a challenging task, especially when balancing performance across high- and low-resource languages. Recent advances in sequence modeling suggest that architectures beyond Transformers may offer better scalability and efficiency. In this work, we introduce MLMA (Multilingual Language Modeling with Mamba for ASR), a new approach that leverages the Mamba architecture -- an efficient state-space model optimized for long-context sequence processing -- for multilingual ASR. Using Mamba, MLMA implicitly incorporates language-aware conditioning and shared representations to support robust recognition across diverse languages. Experiments on standard multilingual benchmarks show that MLMA achieves competitive performance compared to Transformer-based architectures. These results highlight Mamba's potential as a strong backbone for scalable, efficient, and accurate multilingual speech recognition.

Abdul Hannan, Daniele Falavigna, Alessio Brutti

Curating foundation speech models for edge and IoT settings, where computational resources vary over time, requires dynamic architectures featuring adaptable reduction strategies. One emerging approach is layer dropping ($\mathcal{LD}$) which skips fraction of the layers of a backbone network during inference to reduce the computational load. This allows transforming static models into dynamic ones. However, existing approaches exhibit limitations either in the mode of selecting layers or by significantly modifying the neural architecture. To this end, we propose input-driven $\mathcal{LD}$ that employs the network's input features and a lightweight layer selecting network to determine the optimum combination of processing layers. Extensive experimentation on 4 speech and audio public benchmarks, using two different pre-trained foundation models, demonstrates the effectiveness of our approach, thoroughly outperforming random dropping and producing on-par (or better) results to early exit.

Marco Gaido, Sara Papi, Luisa Bentivogli et al.

Training large-scale models presents challenges not only in terms of resource requirements but also in terms of their convergence. For this reason, the learning rate (LR) is often decreased when the size of a model is increased. Such a simple solution is not enough in the case of speech-to-text (S2T) trainings, where evolved and more complex variants of the Transformer architecture -- e.g., Conformer or Branchformer -- are used in light of their better performance. As a workaround, OWSM designed a double linear warmup of the LR, increasing it to a very small value in the first phase before updating it to a higher value in the second phase. While this solution worked well in practice, it was not compared with alternative solutions, nor was the impact on the final performance of different LR warmup schedules studied. This paper fills this gap, revealing that i) large-scale S2T trainings demand a sub-exponential LR warmup, and ii) a higher LR in the warmup phase accelerates initial convergence, but it does not boost final performance.

Abdul Hannan, Alessio Brutti, Shah Nawaz et al.

Recent advancement in deep learning encouraged developing large automatic speech recognition (ASR) models that achieve promising results while ignoring computational and memory constraints. However, deploying such models on low resource devices is impractical despite of their favorable performance. Existing approaches (pruning, distillation, layer skip etc.) transform the large models into smaller ones at the cost of significant performance degradation or require prolonged training of smaller models for better performance. To address these issues, we introduce an efficacious two-step representation learning based approach capable of producing several small sized models from a single large model ensuring considerably better performance in limited number of epochs. Comprehensive experimentation on ASR benchmarks reveals the efficacy of our approach, achieving three-fold training speed-up and up to 12.54% word error rate improvement.

Umberto Cappellazzo, Muqiao Yang, Daniele Falavigna et al.

The ability to learn new concepts sequentially is a major weakness for modern neural networks, which hinders their use in non-stationary environments. Their propensity to fit the current data distribution to the detriment of the past acquired knowledge leads to the catastrophic forgetting issue. In this work we tackle the problem of Spoken Language Understanding applied to a continual learning setting. We first define a class-incremental scenario for the SLURP dataset. Then, we propose three knowledge distillation (KD) approaches to mitigate forgetting for a sequence-to-sequence transformer model: the first KD method is applied to the encoder output (audio-KD), and the other two work on the decoder output, either directly on the token-level (tok-KD) or on the sequence-level (seq-KD) distributions. We show that the seq-KD substantially improves all the performance metrics, and its combination with the audio-KD further decreases the average WER and enhances the entity prediction metric.

Samuele Cornell, Alessio Brutti, Marco Matassoni et al.

Fully exploiting ad-hoc microphone networks for distant speech recognition is still an open issue. Empirical evidence shows that being able to select the best microphone leads to significant improvements in recognition without any additional effort on front-end processing. Current channel selection techniques either rely on signal, decoder or posterior-based features. Signal-based features are inexpensive to compute but do not always correlate with recognition performance. Instead decoder and posterior-based features exhibit better correlation but require substantial computational resources. In this work, we tackle the channel selection problem by proposing MicRank, a learning to rank framework where a neural network is trained to rank the available channels using directly the recognition performance on the training set. The proposed approach is agnostic with respect to the array geometry and type of recognition back-end. We investigate different learning to rank strategies using a synthetic dataset developed on purpose and the CHiME-6 data. Results show that the proposed approach is able to considerably improve over previous selection techniques, reaching comparable and in some instances better performance than oracle signal-based measures.

Vandana Rajan, Alessio Brutti, Andrea Cavallaro

Multi-modal learning relates information across observation modalities of the same physical phenomenon to leverage complementary information. Most multi-modal machine learning methods require that all the modalities used for training are also available for testing. This is a limitation when the signals from some modalities are unavailable or are severely degraded by noise. To address this limitation, we aim to improve the testing performance of uni-modal systems using multiple modalities during training only. The proposed multi-modal training framework uses cross-modal translation and correlation-based latent space alignment to improve the representations of the weaker modalities. The translation from the weaker to the stronger modality generates a multi-modal intermediate encoding that is representative of both modalities. This encoding is then correlated with the stronger modality representations in a shared latent space. We validate the proposed approach on the AVEC 2016 dataset for continuous emotion recognition and show the effectiveness of the approach that achieves state-of-the-art (uni-modal) performance for weaker modalities.

Gianmarco Cerutti, Rahul Prasad, Alessio Brutti et al.

Outdoor acoustic events detection is an exciting research field but challenged by the need for complex algorithms and deep learning techniques, typically requiring many computational, memory, and energy resources. This challenge discourages IoT implementation, where an efficient use of resources is required. However, current embedded technologies and microcontrollers have increased their capabilities without penalizing energy efficiency. This paper addresses the application of sound event detection at the edge, by optimizing deep learning techniques on resource-constrained embedded platforms for the IoT. The contribution is two-fold: firstly, a two-stage student-teacher approach is presented to make state-of-the-art neural networks for sound event detection fit on current microcontrollers; secondly, we test our approach on an ARM Cortex M4, particularly focusing on issues related to 8-bits quantization. Our embedded implementation can achieve 68% accuracy in recognition on Urbansound8k, not far from state-of-the-art performance, with an inference time of 125 ms for each second of the audio stream, and power consumption of 5.5 mW in just 34.3 kB of RAM.

Md Sahidullah, Jose Patino, Samuele Cornell et al.

This paper describes the speaker diarization systems developed for the Second DIHARD Speech Diarization Challenge (DIHARD II) by the Speed team. Besides describing the system, which considerably outperformed the challenge baselines, we also focus on the lessons learned from numerous approaches that we tried for single and multi-channel systems. We present several components of our diarization system, including categorization of domains, speech enhancement, speech activity detection, speaker embeddings, clustering methods, resegmentation, and system fusion. We analyze and discuss the effect of each such component on the overall diarization performance within the realistic settings of the challenge.

Enrico Fini, Alessio Brutti

Recently, a fully supervised speaker diarization approach was proposed (UIS-RNN) which models speakers using multiple instances of a parameter-sharing recurrent neural network. In this paper we propose qualitative modifications to the model that significantly improve the learning efficiency and the overall diarization performance. In particular, we introduce a novel loss function, we called Sample Mean Loss and we present a better modelling of the speaker turn behaviour, by devising an analytical expression to compute the probability of a new speaker joining the conversation. In addition, we demonstrate that our model can be trained on fixed-length speech segments, removing the need for speaker change information in inference. Using x-vectors as input features, we evaluate our proposed approach on the multi-domain dataset employed in the DIHARD II challenge: our online method improves with respect to the original UIS-RNN and achieves similar performance to an offline agglomerative clustering baseline using PLDA scoring.

Xinyuan Qian, Andrea Cavallaro, Alessio Brutti et al.

This document describes our submission to the 2018 LOCalization And TrAcking (LOCATA) challenge (Tasks 1, 3, 5). We estimate the 3D position of a speaker using the Global Coherence Field (GCF) computed from multiple microphone pairs of a DICIT planar array. One of the main challenges when using such an array with omnidirectional microphones is the front-back ambiguity, which is particularly evident in Task 5. We address this challenge by post-processing the peaks of the GCF and exploiting the attenuation introduced by the frame of the array. Moreover, the intermittent nature of speech and the changing orientation of the speaker make localization difficult. For Tasks 3 and 5, we also employ a Particle Filter (PF) that favors the spatio-temporal continuity of the localization results.