Speech Emotion Recognition in Acted and Spontaneous Context

Abstract

Little attention has been paid so far in the context in which databases used for the study of emotion through vocal channel are recorded. Thus, we propose and evaluate an emotion classification system focusing on the differences between acted and spontaneous emotional speech through the use of two different databases: SAVEE and IEMOCAP. For the purpose of this work, we have examined wavelet packet energy and entropy features applied to Mel, Bark and ERB scale applied with Hidden Markov Model (HMM) as classification system. Experimental results show that the proposed method is a feasible technique for emotion classification for both acted and spontaneous context, pointing out the performance difference of the system between the two contexts. The experimental results shows that ERB scale features gives better performance in comparison with other studied features with recognition accuracy of 78.75% for acted context and 50.06% for spontaneous context.

JVNV: A Corpus of Japanese Emotional Speech With Verbal Content and Nonverbal Expressions

“…As mentioned, the number of emotional corpora with nonverbal expressions is limited. Emotional speech corpora can be roughly categorized into two types: acted and spontaneous [8]. In acted corpora, scripts and a recording instruction are provided to the speakers before the recording, which makes it easy to control label balance and almost does not need extra manual annotation.…”

“…To show the contributions of NVs on emotion recognizability, we also evaluate the verbal parts of JVNV by removing nonverbal parts from the utterances of JVNV, which is denoted as ''JVNV-V''. We conducted a forced choice task on a Japanese crowdsourcing platform 8 . For each corpus, we randomly picked up 60 emotion-balanced samples.…”

Few-shot Learning in Emotion Recognition of Spontaneous Speech Using a Siamese Neural Network with Adaptive Sample Pair Formation

“…Due to these limitations, audio samples in many speechbased emotion datasets are collected through acted elicitation methods relying on individuals who engender a target emotion while uttering pre-determined linguistic contents, also known as scripted speech [6]. Despite the fact that these methods tend to overlook subtle expression details, they provide ample data, based on which machine learning (ML) methodologies can recognize emotions [7].…”