Detection of Speech Emotions Using Deep Learning Techniques

Emotion recognition from speech signals is crucial because the need for recognizing different emotions through speech is increasing day by day. Speech Emotion Recognition (SER) is the endeavor to discern human affective and emotional states from speech. Numerous strategies have been used to extract emotions from signals, including various well-established feature extraction and classification techniques of which Deep Learning techniques are gaining prominence. Many studies show that mental healthcare often requires gender-specific treatments. The existing SER systems do not incorporate gender information about the speaker and hence cannot be applied in telemedicine, especially for mental health care services where gender-specific treatments are necessary. This paper focuses on the relevance of gender information incorporated into SER systems and their applications in mental health care services. In this paper, Convolutional Neural Network (CNN) and CNN with Long Short-Term Memory (LSTM) are the models used, and the results are compared for choosing the best model. From the results, it is observed that SER using CNN-LSTM has high accuracy compared to the CNN model. Gender information is incorporated in both CNN and LSTM models to check if it contributes to the increase or decrease in the accuracy of the models. The dataset used is RAVDESS. A web application for SER using CNN-LSTM is implemented considering six emotion categories. This paper also includes a detailed review of the dataset, feature extraction methods, augmentation of the dataset, preparation of the dataset, and graphs associated with it.