Daniele Falavigna

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.

George August Wright, Umberto Cappellazzo, Salah Zaiem et al.

The ability to dynamically adjust the computational load of neural models during inference is crucial for on-device processing scenarios characterised by limited and time-varying computational resources. A promising solution is presented by early-exit architectures, in which additional exit branches are appended to intermediate layers of the encoder. In self-attention models for automatic speech recognition (ASR), early-exit architectures enable the development of dynamic models capable of adapting their size and architecture to varying levels of computational resources and ASR performance demands. Previous research on early-exiting ASR models has relied on pre-trained self-supervised models, fine-tuned with an early-exit loss. In this paper, we undertake an experimental comparison between fine-tuning pre-trained backbones and training models from scratch with the early-exiting objective. Experiments conducted on public datasets reveal that early-exit models trained from scratch not only preserve performance when using fewer encoder layers but also exhibit enhanced task accuracy compared to single-exit or pre-trained models. Furthermore, we explore an exit selection strategy grounded in posterior probabilities as an alternative to the conventional frame-based entropy approach. Results provide insights into the training dynamics of early-exit architectures for ASR models, particularly the efficacy of training strategies and exit selection methods.

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.

Abdul Hannan, Daniele Falavigna, Shah Nawaz et al.

Edge devices operate in constrained and varying resource settings, requiring dynamic architectures that can adapt to limitations of the available resources. To meet such demands, layer dropping ($\mathcal{LD}$) approach is typically used to transform static models into dynamic ones by skipping parts of the network along with reducing overall computational complexity. However, existing $\mathcal{LD}$ methods greatly impact the dynamic model's performance for low and high dropping cases, deteriorating the performance-computation trade-off. To this end, we propose a distillation-based layer dropping (DLD) framework that effectively combines the capabilities of knowledge distillation and $\mathcal{LD}$ in an end-to-end fashion, thereby achieving state-of-the-art performance for dynamic speech networks. Comprehensive experimentation utilizing well-known speech recognition methods, including conformer and WavLM, on three public benchmarks demonstrates the effectiveness of our framework, reducing the word error rate by $9.32\%$ and $2.25\%$ for high and no dropping cases with $33.3\%$ reduction in training time.

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.

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.

Maxence Lasbordes, Daniele Falavigna, Alessio Brutti

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.

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.

Roberto Gretter, Marco Matassoni, Daniele Falavigna

We address the problem of language model customization in applications where the ASR component needs to manage domain-specific terminology; although current state-of-the-art speech recognition technology provides excellent results for generic domains, the adaptation to specialized dictionaries or glossaries is still an open issue. In this work we present an approach for automatically selecting sentences, from a text corpus, that match, both semantically and morphologically, a glossary of terms (words or composite words) furnished by the user. The final goal is to rapidly adapt the language model of an hybrid ASR system with a limited amount of in-domain text data in order to successfully cope with the linguistic domain at hand; the vocabulary of the baseline model is expanded and tailored, reducing the resulting OOV rate. Data selection strategies based on shallow morphological seeds and semantic similarity viaword2vec are introduced and discussed; the experimental setting consists in a simultaneous interpreting scenario, where ASRs in three languages are designed to recognize the domain-specific terms (i.e. dentistry). Results using different metrics (OOV rate, WER, precision and recall) show the effectiveness of the proposed techniques.

Sara Papi, Edmondo Trentin, Roberto Gretter et al.

The paper copes with the task of automatic assessment of second language proficiency from the language learners' spoken responses to test prompts. The task has significant relevance to the field of computer assisted language learning. The approach presented in the paper relies on two separate modules: (1) an automatic speech recognition system that yields text transcripts of the spoken interactions involved, and (2) a multiple classifier system based on deep learners that ranks the transcripts into proficiency classes. Different deep neural network architectures (both feed-forward and recurrent) are specialized over diverse representations of the texts in terms of: a reference grammar, the outcome of probabilistic language models, several word embeddings, and two bag-of-word models. Combination of the individual classifiers is realized either via a probabilistic pseudo-joint model, or via a neural mixture of experts. Using the data of the third Spoken CALL Shared Task challenge, the highest values to date were obtained in terms of three popular evaluation metrics.

Roberto Gretter, Marco Matassoni, Stefano Bannò et al.

This paper describes "TLT-school" a corpus of speech utterances collected in schools of northern Italy for assessing the performance of students learning both English and German. The corpus was recorded in the years 2017 and 2018 from students aged between nine and sixteen years, attending primary, middle and high school. All utterances have been scored, in terms of some predefined proficiency indicators, by human experts. In addition, most of utterances recorded in 2017 have been manually transcribed carefully. Guidelines and procedures used for manual transcriptions of utterances will be described in detail, as well as results achieved by means of an automatic speech recognition system developed by us. Part of the corpus is going to be freely distributed to scientific community particularly interested both in non-native speech recognition and automatic assessment of second language proficiency.

Roberto Gretter, Katharina Allgaier, Svetlana Tchistiakova et al.

This paper describes technology developed to automatically grade Italian students (ages 9-16) on their English and German spoken language proficiency. The students' spoken answers are first transcribed by an automatic speech recognition (ASR) system and then scored using a feedforward neural network (NN) that processes features extracted from the automatic transcriptions. In-domain acoustic models, employing deep neural networks (DNNs), are derived by adapting the parameters of an original out of domain DNN.

Marco Matassoni, Roberto Gretter, Daniele Falavigna et al.

This work deals with non-native children's speech and investigates both multi-task and transfer learning approaches to adapt a multi-language Deep Neural Network (DNN) to speakers, specifically children, learning a foreign language. The application scenario is characterized by young students learning English and German and reading sentences in these second-languages, as well as in their mother language. The paper analyzes and discusses techniques for training effective DNN-based acoustic models starting from children native speech and performing adaptation with limited non-native audio material. A multi-lingual model is adopted as baseline, where a common phonetic lexicon, defined in terms of the units of the International Phonetic Alphabet (IPA), is shared across the three languages at hand (Italian, German and English); DNN adaptation methods based on transfer learning are evaluated on significant non-native evaluation sets. Results show that the resulting non-native models allow a significant improvement with respect to a mono-lingual system adapted to speakers of the target language.

Shahab Jalalvand, Matteo Negri, Daniele Falavigna et al.

Recognizer Output Voting Error Reduction (ROVER) has been widely used for system combination in automatic speech recognition (ASR). In order to select the most appropriate words to insert at each position in the output transcriptions, some ROVER extensions rely on critical information such as confidence scores and other ASR decoder features. This information, which is not always available, highly depends on the decoding process and sometimes tends to over estimate the real quality of the recognized words. In this paper we propose a novel variant of ROVER that takes advantage of ASR quality estimation (QE) for ranking the transcriptions at "segment level" instead of: i) relying on confidence scores, or ii) feeding ROVER with randomly ordered hypotheses. We first introduce an effective set of features to compensate for the absence of ASR decoder information. Then, we apply QE techniques to perform accurate hypothesis ranking at segment-level before starting the fusion process. The evaluation is carried out on two different tasks, in which we respectively combine hypotheses coming from independent ASR systems and multi-microphone recordings. In both tasks, it is assumed that the ASR decoder information is not available. The proposed approach significantly outperforms standard ROVER and it is competitive with two strong oracles that e xploit prior knowledge about the real quality of the hypotheses to be combined. Compared to standard ROVER, the abs olute WER improvements in the two evaluation scenarios range from 0.5% to 7.3%.

Daniele Falavigna, Marco Matassoni, Shahab Jalalvand et al.

In this paper we propose to exploit the automatic Quality Estimation (QE) of ASR hypotheses to perform the unsupervised adaptation of a deep neural network modeling acoustic probabilities. Our hypothesis is that significant improvements can be achieved by: i)automatically transcribing the evaluation data we are currently trying to recognise, and ii) selecting from it a subset of "good quality" instances based on the word error rate (WER) scores predicted by a QE component. To validate this hypothesis, we run several experiments on the evaluation data sets released for the CHiME-3 challenge. First, we operate in oracle conditions in which manual transcriptions of the evaluation data are available, thus allowing us to compute the "true" sentence WER. In this scenario, we perform the adaptation with variable amounts of data, which are characterised by different levels of quality. Then, we move to realistic conditions in which the manual transcriptions of the evaluation data are not available. In this case, the adaptation is performed on data selected according to the WER scores "predicted" by a QE component. Our results indicate that: i) QE predictions allow us to closely approximate the adaptation results obtained in oracle conditions, and ii) the overall ASR performance based on the proposed QE-driven adaptation method is significantly better than the strong, most recent, CHiME-3 baseline.