An Experimental Study on Speech Emotion Recognition for Bangla Language
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Speech emotion recognition (SER) is one of the building blocks of speechenabled personalized applications. However, recognition of emotion from speech is complex due to the variation of emotion intensity (arousal) of the human voice. Data is another challenge for low-resource languages for example, Bangla. To the best of our knowledge, only two datasets are available for SER in Bangla. One of the datasets has not been evaluated using deep learning techniques yet. In this work, we have applied deep learning techniques to available Bangla SER datasets (SUBESCO and BanglaSER) and evaluated their performance. Our major objective is to assess both datasets using identical deep learning methodologies to investigate their usability, a task that has not been undertaken thus far. Convolutional neural networks have been applied to a set of 11 different features of audio data. We have performed extensive experimental study using three different approaches (without cross-validation, cross-validation, and cross-validation with regularization) on three different setups of datasets (combined SUBESCO and Bangla SER, only SUBESCO, and only BanglaSER). Without crossvalidation, more than 70% accuracy has been achieved for all dataset setups. After cross-validation, 68.08% average accuracy and after regularization with cross-validation, 62.62% average accuracy have been achieved on dataset set-up 3 (only BangleSER). The results significantly demonstrate that these datasets are not prepared to develop Bangla SER. Dataset volume is a major challenge for achieving benchmark performance. Hence, the development of a large audio corpus with transcription variation is one of the significant scopes of research in Bangla speech emotion recognition.
Fields of Study & Subjects
Emotion and Mood Recognition
Speech Recognition and Synthesis
Speech and Audio Processing
Publication details
- Journal / source
- CRC Press eBooks
- Pages
- 30–44
- Publication date
- 27 March 2026
- Language
- EN
- Document type
- Book chapter