Emotion Detection from Urdu Speech

This paper addresses the limited existing work for emotion detection from Urdu audio, which has applications for 230 million Urdu speakers for various real-world use cases such as mental health services, AI chatbots, and automated systems such as robocalls. We used a dataset of 13,920 audio snippets in four emotions (anger, happiness, neutral, and sadness) to evaluate several models. The highest accuracy in emotion detection was achieved by the transformer-based, Self-Supervised AST (Audio Spectrogram Transformer) model at an accuracy of 98.71%, outperforming the other seven models we trained as well as those presented in other comparative studies.