Multiscale kernel locally penalised discriminant analysis exemplified by emotion recognition in speech

Abstract

We propose a novel method to learn multiscale kernels with locally penalised discriminant analysis, namely Multiscale-Kernel Locally Penalised Discriminant Analysis (MS-KLPD A). As an exemplary use-case, we apply it to recognise emotions in speech. Specifically, we employ the term of locally penalised discriminant analysis by controlling the weights of marginal sample pairs, while the method learns kernels with multiple scales. Evaluated in a series of experiments on emotional speech corpora, our proposed MS-KLPDA is able to outperform the previous research of Multiscale-Kernel Fisher Discriminant Analysis and some conventional methods in solving speech emotion recognition. CCS Concepts•Information systems → Multimedia information systems; •Human-centered computing → HCI theory, concepts and models; •Theory of computation → Kernel methods;

Connecting Subspace Learning and Extreme Learning Machine in Speech Emotion Recognition

“…Specifically, the conventional linear subspace learning methods PCA, LPP [15], LDA, LDE [16] were considered. The recently proposed linear subspace learning methods, LDP [14], GbFA [17], and LPDA [8], were also used for comparison purposes. Furthermore,we tested three kernel subspace learning methods [10] using LDE, GbFA, and LPDA embedding graphs respectively.…”

“…It should be noticed that we only consider the fully supervised case, where each training sample is labelled by a single emotional class. The embedding graphs (G (I ) and G (P ) ) are obtained using preexistent GE based algorithms -FDA [6], [10], LDP [14], LDE [16], GbFA [17], and Locally Penalised Discriminant Analysis (LPDA; [8]). The description of the FDA and LDP embedding graphs can be found in [6], [10], [14] (also introduced in Section II-B).…”

“…This need is apparent in many human-computer interaction tasks such as Speech Emotion Recognition (SER), which aims to detect emotional information conveyed by the human voice [2]- [5], but many of the typically used acoustic features contain information that may not be relevant for emotion recognition [6], [7]. Thus SER requires new methods for obtaining factors specifically related to emotional representation from large feature spaces [6], [8].…”