Speech Recognition Using Hidden Markov Model

Abstract

Automatic speech recognition (ASR) is achieved through various models among which, Hidden Markov Model (HMM) gives optimum probabilities for speech extraction. The power of the HMM lies in dividing the signal probability function into smaller divisions (thus overcoming the condition of stationarity). Also the recursive nature of the model gives better convergence, therefore more accurate results. This thesis focuses on use of HMM for automatic speech recognition. Towards this purpose MATLAB was used to simulate the HMM behaviour. The features codebook was build from the speech files of the isolated words YES, NO, CAT were recorded for many times from the same person and recognized the 10 (YES, NO, CAT) words spoken by the same person as hidden Markov model’s output.