Analysis of the effectiveness of ML algorithms for emotion recognition, taking into account prosodic and spectral features

Abstract

The aim of the study is to determine the optimal classifier for identifying an emotional state based on the results of a comparative analysis of the effectiveness of various machine learning algorithms based on a combination of prosodic and spectral features. The scientific novelty consists in the application of ML algorithms in the recognition of emotionally marked speech of North Caucasian bilinguals in the problem of binary classification of the presence or absence of an accent with the determination of the optimal combination of universal prosodic and spectral features. During the study, an experimental corpus of speech of representatives of three ethnic groups (Russians, Kabardians and Armenians) was created with an annotation of the degree of accent, prosodic (94 signs) and spectral (74 signs) characteristics were extracted from speech signals, a comparative analysis of the effectiveness of machine learning algorithms (logistic regression, k-nearest neighbors, the method of support vectors, decision trees) in the problem of binary classification of the presence/absence of emphasis. The results of the study showed that at the syllabic level, the most effective is the decision tree model with combined features, and at the phrasal level, the k-nearest neighbor model with prosodic features. Universal prosodic features that form the basis of the "language model of emotions" were identified, as well as typological differences in their implementation, reflecting the influence of the native language on the emotional speech of bilinguals.