Zero Crossing Rate (ZCR) Feature Extraction in Speech Emotion Recognition Using a Convolutional Neural Network

Abstract

Speech Emotion Recognition (SER) plays a pivotal role in enhancing human-computer interaction (HCI), healthcare, and automated feedback systems (Hashem et al., 2023). However, extracting relevant acoustic features from complex speech signals remains a significant challenge (Lieskovská et al., 2021). This paper investigates the effectiveness of the Zero Crossing Rate (ZCR) feature extraction method combined with a 1D Convolutional Neural Network (CNN) architecture for speech emotion classification. By evaluating emotional speech data across benchmark datasets (RAVDESS, CREMA, SAVEE, and TESS), audio signals were preprocessed using data augmentation techniques, including noise injection, time stretching, pitch shifting, and time shifting. The ZCR feature representation was fed into a CNN model to classify seven core emotional states: anger, disgust, fear, happiness, sadness, surprise, and neutral. The proposed ZCR-CNN approach achieved a classification accuracy of 61.96% with a training loss of 0.8214 and testing loss of 0.8123. Although ZCR effectively captures high-frequency and noisiness variations indicative of specific emotional states, its performance highlights the trade-offs of relying on single-domain time features compared to spectral features (Singh & Goel, 2022).