MNITJ-SEHSD: A Hindi Emotional Speech Database

This paper introduces an emotional speech database in Hindi to study the emotions embedded in speech signals for speech emotion recognition. The proposed database was created with the help of students from Malaviya National Institute of Technology, Jaipur, India. It is compiled employing five distinct emotions using some emotionless texts. Anger, Fear, Happy, Neutral, and Sad are the emotions presented in the database, considering these are the most common emotions in daily life. The speech corpus is developed at MNIT, Jaipur, named Malaviya National Institute of Technology Jaipur Simulated Emotion Hindi Speech Database (MNITJ-SEHSD). The built database’s emotion classification used various well-known audio features. Mel frequency cepstral coefficients (MFCCs), chromagram, and Mel-spectrogram extract spectral information. Prosody information is represented by pitch, the standard deviation of pitch, energy, duration and zero-crossing rate. The performance of emotion recognition using the aforementioned audio representations is around 57% and 90% for prosodic and spectral features, respectively. The database is also validated on our previously developed CNN model. The development, collection, processing, and evaluation of the proposed speech database are all described in this work. Subjective listening methods evaluate the overall quality of the emotions stored in the database.