Hybrid Models for Speech Emotion Detection: A Fusion of Traditional and Deep Learning Techniques

Emotion is a vital aspect of human life, characterized by a multifaceted psychological and physiological state encompassing a range of thoughts, feelings, and physical reactions. This research attempts to extract real emotions from speech, addressing the common problem of accurately detecting emotions due to their varying intensity and duration using the RAVDESS dataset. It focuses on multiple emotions, varying intensity and length, such as happiness, sadness, fear, anger, surprise, and disgust. In the preprocessing stage, detailed waveform and spectrogram analysis, and One-hot coding were used to assign seven different emotion labels. Important characteristics, including pitch, magnitude, chroma, Mel-Frequency Cepstral Coefficients (MFCC) and pre-trained VGGish and OpenL3 were retrieved to condense the content effectively. Various machine learning and deep learning models were employed, including Support Vector Machine (SVM), stacking ensemble methods, Convolutional Neural Networks (CNN), Deep Stride Convolutional Neural Networks (DSCNN), Long Short-Term Memory networks (LSTMs). We proposed hybrid models combining the strong models, CNN-SVM and CNN-LSTM. These models were selected based on how well they could extract features, classify data, and capture the temporal correlations in voice data. The study highlights the efficacy of the CNN-SVM hybrid model, which showed exceptional performance with an astounding accuracy rate of 98% in distinguishing genuine emotions from speech. This high accuracy underscores the potential of hybrid models in advancing speech recognition technologies.