Improving Emotion Detection in Audio Using Deep Learning Algorithms

Emotion analysis in audio-based systems has become an important research topic in recent years, as understanding speech-based human emotions has various practical applications, such as improving human-computer interaction, enhancing emotional communication, and aiding in mental health support via audio-based chatbot systems. Emotion detection using audio faces limitations such as emotion ambiguity in audio files, limitations in feature representations, and generalization concerning varying accents and dialects. This paper showcases a comparative analysis of various deep-learning algorithms for emotion analysis via audio signals. Spectrogram-based feature representation and audio-based feature representations have also been implemented and compared. To evaluate the effectiveness of the approach, the authors conducted experiments on the publicly available SAVEE dataset. Among the implemented algorithms, CNN (with spectrograms) and RNN (with MFCC) have showcased an improved accuracy of 61%.