Development of a speech dataset for emotion recognition and analysis

In recent years, with the development of technologies such as signal processing, machine learning, and deep learning, speech emotion recognition has found wide applications in mental health, educational training, and social media analysis. However, current methods for speech emotion recognition mostly rely on data-driven models, and the lack of emotional speech datasets significantly hinders their progress. To address this issue, this paper proposes an emotional speech dataset consisting of 6,447 audio clips for emotion recognition and analysis. The dataset includes five emotional categories: happy, sad, surprised, fearful, and angry. In addition, four speech signal processing algorithms including Fast Fourier Transform (FFT), Short-Time Fourier Transform (STFT), Mel-Frequency Cepstral Coefficients (MFCC), and Continuous Wavelet Transform (CWT) are utilized for feature extraction, resulting in the creation of four emotional feature image datasets. Furthermore, eleven state-of-the-art (SOTA) classification models are selected to establish a benchmark. Experimental results show that most classification models achieve accurate recognition performance on the proposed emotional speech dataset. It is believed that the proposed dataset and benchmark can promote the research and development of emotion recognition and analysis.