View article

Speaker attributed variability are undesirable in speaker independent speech recognition systems. The gender of the speaker is one of the influential sources of this variability. Common speech recognition systems tuned to the ensemble statistics over many speakers to compensate the inherent variability of speech signal. In this paper we will separate the datasets based on the gender to build gender dependent hidden Markov model for each word. The gender separation criterion is the average pitch frequency of the speaker. Experimental evaluation shows significant improvement in word recognition accuracy over the gender independent method with a slight increase in the processing computation.