Design and Validation of HindiSER:Speech Emotion Recognition Dataset for Hindi Language

• Main Authors: Shinde, Ashwini S., Patil, Vaishali V., Shaikh, Albaab, More, Pratik, Salve, Kajal
• Format: Conference Proceeding
• Language: English
• Subjects: Deep learning; Emotion recognition; Hindi Speech Emotion Recognition Database; Human computer interaction; KNN; Low Resource Regional language emotion database; MFCC; MLP; Sociology; Spectral features; Support vector machines; SVM; Tongue
• ISSN: 2472-7695
• DOI: 10.1109/O-COCOSDA60357.2023.10482916

Every human prefers to express their emotions in their mother tongue. In India, Hindi is considered the National Language and is spoken by around 45% of the population. The unavailability of emotional speech databases in many of the regional languages is a significant bottleneck in the domain of Human-Computer Interaction (HCI). To address this issue of Low-resource regional language, we have proposed a new Hindi language emotional speech database "HindiSER". This database has a total 520 emotional utterances recorded from 26 non-professional native Hindi speakers in the age group of 19-45 years. HindiSER has emotional utterances of happy, sad, neutral, and angry emotion classes of 5 emotionally neutral Hindi sentences. Thus, the number of utterances in each emotional class is 130. Objective validation is performed with Support Vector Machine (SVM) classifier, K-Nearest Neighbor (KNN) and also with Multi-Layer Perceptron (MLP) classifier. Subjective validation of HindiSER is done through a listening test. Performance comparison of these classifiers using spectral features shows better recall on the happy emotion class and lowest recall on neutral emotion.