Parkinson’s Disease (PD) is one of the most common non-curable neurodegenerative diseases. Diagnosis is achieved clinically on the basis of different symptoms with considerable delays from the onset of neurodegenerative processes in the central nervous system. In this study, we investigated early and full-blown PD patients based on the analysis of their voice characteristics with the aid of the most commonly employed machine learning (ML) techniques. A custom dataset was made with hi-fi quality recordings of vocal tasks gathered from Italian healthy control subjects and PD patients, divided into early diagnosed, off-medication patients on the one hand, and mid-advanced patients treated with L-Dopa on the other. Following the current state-of-the-art, several ML pipelines were compared usingdifferent feature selection and classification algorithms, and deep learning was also explored with a custom CNN architecture. Results show how feature-based ML and deep learning achieve comparable results in terms of classification, with KNN, SVM and naïve Bayes classifiers performing similarly, with a slight edge for KNN. Much more evident is the predominance of CFS as the best feature selector. The selected features act as relevant vocal biomarkers capable of differentiating healthy subjects, early untreated PD patients and mid-advanced L-Dopa treated patients.

Machine Learning (ML) algorithms within a human–computer framework are the leading force in speech emotion recognition (SER). However, few studies explore cross-corpora aspects of SER; this work aims to explore the feasibility and characteristics of a cross-linguistic, cross-gender SER. Three ML classifiers (SVM, Naïve Bayes and MLP) are applied to acoustic features, obtained through a procedure based on Kononenko’s discretization and correlation-based feature selection. The system encompasses five emotions (disgust, fear, happiness, anger and sadness), using the Emofilm database, comprised of short clips of English movies and the respective Italian and Spanish dubbed versions, for a total of 1115 annotated utterances. The results see MLP as the most effective classifier, with accuracies higher than 90% for single-language approaches, while the cross-language classifier still yields accuracies higher than 80%. The results show cross-gender tasks to be more difficult than those involving two languages, suggesting greater differences between emotions expressed by male versus female subjects than between different languages. Four feature domains, namely, RASTA, F0, MFCC and spectral energy, are algorithmically assessed as the most effective, refining existing literature and approaches based on standard sets. To our knowledge, this is one of the first studies encompassing cross-gender and cross-linguistic assessments on SER.

Speaker Recognition (SR) is a common task in AI-based sound analysis, involving structurally different methodologies such as Deep Learning or “traditional” Machine Learning (ML). In this paper, we compared and explored the two methodologies on the DEMoS dataset consisting of 8869 audio files of 58 speakers in different emotional states. A custom CNN is compared to several pre-trained nets using image inputs of spectrograms and Cepstral-temporal (MFCC) graphs. AML approach based on acoustic feature extraction, selection and multi-class classification by means of a Naïve Bayes model is also considered. Results show how a custom, less deep CNN trained on grayscale spectrogram images obtain the most accurate results, 90.15% on grayscale spectrograms and 83.17% on colored MFCC. AlexNet provides comparable results, reaching 89.28% on spectrograms and 83.43% on MFCC.The Naïve Bayes classifier provides a 87.09% accuracy and a 0.985 average AUC while being faster to train and more interpretable. Feature selection shows how F0, MFCC and voicing-related features are the most characterizing for this SR task. The high amount of training samples and the emotional content of the DEMoS dataset better reflect a real case scenario for speaker recognition, and account for the generalization power of the models.

Parkinson’s Disease and Adductor-type Spasmodic Dysphonia are two neurological disorders that greatly decrease the quality of life of millions of patients worldwide. Despite this great diffusion, the related diagnoses are often performed empirically, while it could be relevant to count on objective measurable biomarkers, among which researchers have been considering features related to voice impairment that can be useful indicators but that can sometimes lead to confusion. Therefore, here, our purpose was aimed at developing a robust Machine Learning approach for multi-class classification based on 6373 voice features extracted from a convenient voice dataset made of the sustained vowel/e/ and an ad hoc selected Italian sentence, performed by 111 healthy subjects, 51 Parkinson’s disease patients, and 60 dysphonic patients. Correlation, Information Gain, Gain Ratio, and Genetic Algorithm-based methodologies were compared for feature selection, to build subsets analyzed by means of Naïve Bayes, Random Forest, and Multi-Layer Perceptron classifiers, trained with a 10-fold cross-validation. As a result, spectral, cepstral, prosodic, and voicing-related features were assessed as the most relevant, the Genetic Algorithm performed as the most effective feature selector, while the adopted classifiers performed similarly. In particular, a Genetic Algorithm + Naïve Bayes approach brought one of the highest accuracies in multi-class voice analysis, being 95.70% for a sustained vowel and 99.46% for a sentence.