Speech Processing for Hindi Dialect Recognition

In this paper, the authors have used 2-layer feed forward neural network for Hindi dialect recognition. A Dialect is a pattern of pronunciation of a language used by a community of native speakers belonging to the same geographical region. In this work, speech features have been explored to recognize four major dialects of Hindi. The dialects under consideration are Khariboli (spoken in West Uttar Pradesh, Delhi and some parts of Uttarakhand and Himachal Pradesh), Bhojpuri (spoken by population of East Uttar Pradesh, Bihar and Jharkhand), Haryanvi (spoken in Haryana, parts of Delhi, Uttar Pradesh and Uttarakhand) and Bagheli (spoken in Central India). Speech corpus for this work is collected from 15 speakers (including both male and female) from each dialect. The syllables of CVC structure is used as processing unit. Spectral features (MFCC) and prosodic features (duration and pitch contour) are extracted from speech for discriminating the dialects. Performance of the system is observed with spectral features and prosodic features as input. Results show that the system performs best when all the spectral and prosodic features are combined together to form input feature set during network training. The dialect recognition system shows a recognition score of 79% with these input features.