Transfer learning for Speech Based Emotion Recognition

Speech emotion recognition has an important role in computer-mediated human communication. One of the major challenges in speech-based emotion recognition is data scarcity. Emotions of spontaneous speech in real scenarios and non-labeled speech data emotions are difficult to identify. Transfer learning is an effective way to address challenges related to data scarcity and the lack of human labels. This work uses the transfer learning method based Time Delay Neural Network (TDNN), which uses the knowledge learned from one domain and applies it to another domain with fewer data. Malayalam is one of the less-resourced languages in the case of emotional speech data. Hence the emotional recognition from this is difficult. Transfer learning can be used to identify emotions in languages with non-labeled samples due to the similarity in features corresponding to the emotions. The main reason behind such a technique is that people may share similar characteristics when expressing a given emotion. TDNN was first trained in the source domain to retain and reuse that knowledge to identify the emotions in the target domain. In this work, TDNN uses the RAVDESS English dataset as a source and the Malayalam Speech Corpus in Indic TTS Malayalam dataset as a target to identify the emotions in the Malayalam dataset. The performance of the TDNN based system was compared with a Multilayer Perceptron (MLP) network to study the performance enhancement due to transfer learning. The experimental results show that the accuracy of detection has been enhanced from 57.67% to 88.89%.