Automatic Gender Identification for Hindi Speech Recognition

This paper presents the preliminary work done towards the development of a Gender Recognition System that can be incorporated into the Hindi Automatic Speech Recognition (ASR) System. Gender Recognition (GR) can help in the development of speaker-independent speech recognition systems. This paper presents a general approach to identifying feature vectors that effectively distinguish gender of a speaker from Hindi phoneme utterances. 10 vowels and 5 nasals of the Hindi language were studied for their effectiveness in identifying gender of the speaker. All the 10 vowel Phonemes performed well, while b] bZ] Å] ,] ,s] vks and vkS showed excellent gender distinction performance. All five nasals 3] ́] .k] u and e which were tested, showed a recognition accuracy of almost 100%. The Mel Frequency Cepstral Coefficients (MFCC) are widely used in ASR. The choice of MFCC as features in Gender Recognition will avoid additional computation. The effect of the MFCC feature vector dimension on the GR accuracy was studied and the findings presented. General Terms Automatic speech recognition in Hindi