Many speech features and models, including Deep Neural Networks (DNN), are used for classification tasks between healthy and pathological speech with the Saarbruecken Voice Database (SVD). However, accuracy values of 80.71% for phrases or 82.8% for vowels /aiu/ are the highest reported for audio samples in SVD when the evaluation includes the wide amount of pathologies in the database, instead of a selection of some pathologies. This paper targets this top performance in the state-of-the-art Automatic Voice Disorder Detection (AVDD) systems. In the framework of a DNN-based AVDD system we study the capability of Self-Supervised (SS) representation learning for describing discriminative cues between healthy and pathological speech. The system processes the SS temporal sequence of features with a single feed-forward layer and Class-Token (CT) Transformer for obtaining the classification between healthy and pathological speech. Furthermore, there is evaluated a suitable data extension of the training set with out-of-domain data is also evaluated to deal with the low availability of data for using DNN-based models in voice pathology detection. Experimental results using audio samples corresponding to phrases in the SVD dataset, including all pathologies available, show classification accuracy values until 93.36%. This means that the proposed AVDD system achieved accuracy improvements of 4.1% without the training data extension, and 15.62% after the training data extension compared to the baseline system. Beyond the novelty of using SS representations for AVDD, the fact of obtaining accuracies over 90% in these conditions and using the whole set of pathologies in the SVD is a milestone for voice disorder-related research. Furthermore, the study on the amount of in-domain data in the training set related to the system performance show guidance for the data preparation stage. Lessons learned in this work suggest guidelines for taking advantage of DNN, to boost the performance in developing automatic systems for diagnosis, treatment, and monitoring of voice pathologies.

Speech Emotion Recognition (SER) plays a crucial role in applications involving human-machine interaction. However, the scarcity of suitable emotional speech datasets presents a major challenge for accurate SER systems. Deep Neural Network (DNN)-based solutions currently in use require substantial labelled data for successful training. Previous studies have proposed strategies to expand the training set in this framework by leveraging available emotion speech corpora. This paper assesses the impact of a cross-corpus training extension for a SER system using self-supervised (SS) representations, namely HuBERT and WavLM. The feasibility of training systems with just a few minutes of in-domain audio is also analyzed. The experimental results demonstrate that augmenting the training set with EmoDB (German), RAVDESS, and CREMA-D (English) datasets leads to improved SER accuracy on the IEMOCAP dataset. By combining a cross-corpus training extension and SS representations, state-of-the-art performance is achieved. These findings suggest that the cross-corpus strategy effectively addresses the scarcity of labelled data and enhances the performance of SER systems.

This paper proposes a Deep Learning (DL) based Wiener filter estimator for speech enhancement in the framework of the classical spectral-domain speech estimator algorithm. According to the characteristics of the intermediate steps of the speech enhancement algorithm, i.e., the SNR estimation and the gain function, there is determined the best usage of the network at learning a robust instance of the Wiener filter estimator. Experiments show that the use of data-driven learning of the SNR estimator provides robustness to the statistical-based speech estimator algorithm and achieves performance on the state-of-the-art. Several objective quality metrics show the performance of the speech enhancement and beyond them, there are examples of noisy vs. enhanced speech available for listening to demonstrate in practice the skills of the method in simulated and real audio.

Speech Activity Detection (SAD) aims to accurately classify audio fragments containing human speech. Current state-of-the-art systems for the SAD task are mainly based on deep learning solutions. These applications usually show a significant drop in performance when test data are different from training data due to the domain shift observed. Furthermore, machine learning algorithms require large amounts of labelled data, which may be hard to obtain in real applications. Considering both ideas, in this paper we evaluate three unsupervised domain adaptation techniques applied to the SAD task. A baseline system is trained on a combination of data from different domains and then adapted to a new unseen domain, namely, data from Apollo space missions coming from the Fearless Steps Challenge. Experimental results demonstrate that domain adaptation techniques seeking to minimise the statistical distribution shift provide the most promising results. In particular, Deep CORAL method reports a 13% relative improvement in the original evaluation metric when compared to the unadapted baseline model. Further experiments show that the cascaded application of Deep CORAL and pseudo-labelling techniques can improve even more the results, yielding a significant 24% relative improvement in the evaluation metric when compared to the baseline system.

Metric learning approaches have widely expanded to the training of Speaker Verification (SV) systems based on Deep Neural Networks (DNNs), by using a loss function more consistent with the evaluation process than the traditional identification losses. However, these methods do not consider the performance measure and can involve high computational cost, for example, the need for a careful pair or triplet data selection. This paper proposes the approximated Detection Cost Function (aDCF) loss, which is a loss function based on the measure of the decision errors in SV systems, namely the False Rejection Rate (FRR) and the False Acceptance Rate (FAR). With aDCF loss as the training objective function, the end-to-end system learns how to minimize decision errors. Furthermore, we replace the typical linear layer as the last layer of DNN by a cosine distance layer, which reduces the difference between the metric in the training process and the metric during evaluation. aDCF loss function was evaluated in RSR2015-Part I and RSR2015-Part II datasets for text-dependent speaker verification. The system trained with aDCF loss outperforms all the state-of-the-art functions employed in this paper in both parts of the database.

This paper studies the Speech Enhancement based on Deep Neural Networks. The proposed architecture gradually follows the signal transformation during enhancement by means of a visualization probe at each network block. Alongside the process, the enhancement performance is visually inspected and evaluated in terms of regression cost. This progressive scheme is based on Residual Networks. During the process, we investigate a residual connection with a constant number of channels, including internal state between blocks, and adding progressive supervision. The insights provided by the interpretation of the network enhancement process leads us to design an improved architecture for the enhancement purpose. Following this strategy, we are able to obtain speech enhancement results beyond the state-of-the-art, achieving a favorable trade-off between dereverberation and the amount of spectral distortion.