Abstract While neural networks have demonstrated remarkable predictive capabilities in various scenarios, they typically struggle to learn to avoid regions of the output space that are considered off-limits due to known domain-specific constraints. This paper presents a novel primal-dual learning approach inspired by augmented Lagrangian methods to address a priori output constraints for neural network predictions. Our solution encodes domain constraints for the output space by using a static Implicit Neural Representation and penalising the violation of these constraints at the loss level. This choice allows full flexibility in incorporating the constraints. We conduct extensive evaluations on several 2D and 3D synthetic datasets with different constraint topologies and two real-world datasets to evaluate the effectiveness of the proposed method. Our approach consistently outperforms methods without constraints in all experiments, yielding higher accuracy and lower constraint violations. Finally, our method shows superior performance improvements in scenarios with limited data availability, opening up potential benefits for various applications. These include geo-localisation tasks, where accurate positioning is crucial, and physical problems with theoretically-imposed constraints.

Speech foundation models have demonstrated exceptional capabilities in speech-related tasks. Nevertheless, these models often struggle with non-verbal audio data, such as vocalizations, baby crying, etc., which are critical for various real-world applications. Audio foundation models well handle non-speech data but also fail to capture the nuanced features of non-verbal human sounds. In this work, we aim to overcome the above shortcoming and propose a novel foundation model, termed voc2vec, specifically designed for non-verbal human data leveraging exclusively open-soruce non-verbal audio datasets. We employ a collection of 10 datasets covering around 125 hours of non-verbal audio. Experimental results prove that voc2vec is effective in non-verbal vocalization classification, and it outperforms conventional speech and audio foundation models. Moreover, voc2vec consistently outperforms strong baselines, namely OpenSmile and emotion2vec, on six different benchmark datasets. To the best of the authors’ knowledge, voc2vec is the first universal representation model for vocalization tasks.

Voice disorders significantly impact patient quality of life, yet non-invasive automated diagnosis remains under-explored due to both the scarcity of pathological voice data, and the variability in recording sources. This work introduces MVP (Multi-source Voice Pathology detection), a novel approach that leverages transformers operating directly on raw voice signals. We explore three fusion strategies to combine sentence reading and sustained vowel recordings: waveform concatenation, intermediate feature fusion, and decision-level combination. Empirical validation across the German, Portuguese, and Italian languages shows that intermediate feature fusion using transformers best captures the complementary characteristics of both recording types. Our approach achieves up to +13% AUC improvement over single-source methods.

Kolmogorov-Arnold Networks (KANs) have recently emerged as a promising alternative to traditional neural architectures, yet their application to speech processing remains under explored. This work presents the first investigation of KANs for Spoken Language Understanding (SLU) tasks. We experiment with 2D-CNN models on two datasets, integrating KAN layers in five different configurations within the dense block. The best-performing setup, which places a KAN layer between two linear layers, is directly applied to transformer-based models and evaluated on five SLU datasets with increasing complexity. Our results show that KAN layers can effectively replace the linear layers, achieving comparable or superior performance in most cases. Finally, we provide insights into how KAN and linear layers on top of transformers differently attend to input regions of the raw waveforms.

Despite the remarkable progress in end-to-end Automatic Speech Recognition (ASR) engines, accurately transcribing dysarthric speech remains a major challenge. In this work, we proposed a two-stage framework for the Speech Accessibility Project Challenge at INTERSPEECH 2025, which combines cutting-edge speech recognition models with LLM-based generative error correction (GER). We assess different configurations of model scales and training strategies, incorporating specific hypothesis selection to improve transcription accuracy. Experiments on the Speech Accessibility Project dataset demonstrate the strength of our approach on structured and spontaneous speech, while highlighting challenges in single-word recognition. Through comprehensive analysis, we provide insights into the complementary roles of acoustic and linguistic modeling in dysarthric speech recognition.

Italy exhibits rich linguistic diversity across its territory due to the distinct regional languages spoken in different areas. Recent advances in self-supervised learning provide new opportunities to analyze Italy’s linguistic varieties using speech data alone. This includes the potential to leverage representations learned from large amounts of data to better examine nuances between closely related linguistic varieties. In this study, we focus on automatically identifying the geographic region of origin of speech samples drawn from Italy’s diverse language varieties. We leverage self-supervised learning models to tackle this task and analyze differences and similarities between Italy’s regional languages. In doing so, we also seek to uncover new insights into the relationships among these diverse yet closely related varieties, which may help linguists understand their interconnected evolution and regional development over time and space. To improve the discriminative ability of learned representations, we evaluate several supervised contrastive learning objectives, both as pre-training steps and additional fine-tuning objectives. Experimental evidence shows that pre-trained self-supervised models can effectively identify regions from speech recording. Additionally, incorporating contrastive objectives during fine-tuning improves classification accuracy and yields embeddings that distinctly separate regional varieties, demonstrating the value of combining self-supervised pre-training and contrastive learning for this task.

Eliana Pastor, Alkis Koudounas, Giuseppe Attanasio, Dirk Hovy, Elena Baralis. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

In the domain of music generation, prevailing methods focus on text-to-music tasks, predominantly relying on diffusion models. However, they fail to achieve good vocal quality in synthetic music compositions.To tackle this critical challenge, we present Ainur, a hierarchical diffusion model that concentrates on the lyrics-to-music generation task. Through its use of multimodal Lyrics-Audio Spectrogram Pre-training (CLASP) embeddings, Ainur distinguishes itself from past approaches by specifically enhancing the vocal quality of synthetically produced music. Notably, Ainur’s training and testing processes are highly efficient, requiring only a single GPU. According to experimental results, Ainur meets or exceeds the quality of other state-of-the-art models like MusicGen, MusicLM, and AudioLDM2 in both objective and subjective evaluations. Additionally, Ainur offers near real-time inference speed, which facilitate its use in practical, real-world applications.

As speech processing moves toward more data-hungry models, data selection and acquisition become crucial to building better systems. Recent efforts have championed quantity over quality, following the mantra "The more data, the better." However, not every data brings the same benefit. This paper proposes a data acquisition solution that yields better models with less data – and lower cost. Given a model, a task, and an objective to maximize, we propose a process with three steps. First, we assess the model’s baseline performance on the task. Second, we use efficient mining techniques to identify subgroups that maximize the target objective if acquired first as new samples. Being the subgroups interpretable, we can determine which samples to acquire. Third, we run incremental training sampling from those subgroups. Experiments with two state-of-the-art speech models for Intent Classification across two datasets in English and Italian show that our method is significantly better than random or complete acquisition and clustering-based techniques.

The evaluation of spoken language understanding (SLU) systems is often restricted to assessing their global performance or examining predefined subgroups of interest. However, a more detailed analysis at the subgroup level has the potential to uncover valuable insights into how speech system performance differs across various subgroups. In this work, we identify biased data subgroups and describe them at the level of user demographics, recording conditions, and speech targets. We propose a new task-, model- and dataset-agnostic approach to detect significant intra- and cross-model performance gaps. We detect problematic data subgroups in SLU models by leveraging the notion of subgroup divergence. We also compare the outcome of different SLU models on the same dataset and task at the subgroup level. We identify significant gaps in subgroup performance between models different in size, architecture, or pre-training objectives, including multi-lingual and mono-lingual models, yet comparable to each other in overall performance. The results, obtained on two SLU models, four datasets, and three different tasks–intent classification, automatic speech recognition, and emotion recognition–confirm the effectiveness of the proposed approach in providing a nuanced SLU model assessment.

End-to-End Spoken Language Understanding models are generally evaluated according to their overall accuracy, or separately on (a priori defined) data subgroups of interest. We propose a technique for analyzing model performance at the subgroup level, which considers all subgroups that can be defined via a given set of metadata and are above a specified minimum size. The metadata can represent user characteristics, recording conditions, and speech targets. Our technique is based on advances in model bias analysis, enabling efficient exploration of resulting subgroups. A fine-grained analysis reveals how model performance varies across sub-groups, identifying modeling issues or bias towards specific subgroups.We compare the subgroup-level performance of models based on wav2vec 2.0 and HuBERT on the Fluent Speech Commands dataset. The experimental results illustrate how subgroup-level analysis reveals a finer and more complete picture of performance changes when models are replaced, automatically identifying the subgroups that most benefit or fail to benefit from the change.

Recent large-scale Spoken Language Understanding datasets focus predominantly on English and do not account for language-specific phenomena such as particular phonemes or words in different lects.We introduce ITALIC, the first largescale speech dataset designed for intent classification in Italian.The dataset comprises 16,521 crowdsourced audio samples recorded by 70 speakers from various Italian regions and annotated with intent labels and additional metadata.We explore the versatility of ITALIC by evaluating current state-of-the-art speech and text models.Results on intent classification suggest that increasing scale and running language adaptation yield better speech models, monolingual text models outscore multilingual ones, and that speech recognition on ITALIC is more challenging than on existing Italian benchmarks.We release both the dataset and the annotation scheme to streamline the development of new Italian SLU models and language-specific datasets.

Mosquitoes are a major global health problem. They are responsible for the transmission of diseases and can have a large impact on local economies. Monitoring mosquitoes is therefore helpful in preventing the outbreak of mosquito-borne diseases. In this paper, we propose a novel data-driven approach that leverages Transformer-based models for the identification of mosquitoes in audio recordings. The task aims at detecting the time intervals corresponding to the acoustic mosquito events in an audio signal. We formulate the problem as a sequence tagging task and train a Transformer-based model using a real-world dataset collecting mosquito recordings. By leveraging the sequential nature of mosquito recordings, we formulate the training objective so that the input recordings do not require fine-grained annotations. We show that our approach is able to outperform baseline methods using standard evaluation metrics, albeit suffering from unexpectedly high false negatives detection rates. In view of the achieved results, we propose future directions for the design of more effective mosquito detection models.

Grassmann manifold based sparse spectral clustering is a classification technique that consists in learning a latent representation of data, formed by a subspace basis, which is sparse. In order to learn a latent representation, spectral clustering is formulated in terms of a loss minimization problem over a smooth manifold known as Grassmannian. Such minimization problem cannot be tackled by one of traditional gradient-based learning algorithms, which are only suitable to perform optimization in absence of constraints among parameters. It is, therefore, necessary to develop specific optimization/learning algorithms that are able to look for a local minimum of a loss function under smooth constraints in an efficient way. Such need calls for manifold optimization methods. In this paper, we extend classical gradient-based learning algorithms on at parameter spaces (from classical gradient descent to adaptive momentum) to curved spaces (smooth manifolds) by means of tools from manifold calculus. We compare clustering performances of these methods and known methods from the scientific literature. The obtained results confirm that the proposed learning algorithms prove lighter in computational complexity than existing ones without detriment in clustering efficacy.