Speech Recognition System Based on OLLO French Corpus by Using MFCCs

Abstract

The automatic speech recognition is an area of active study since the early 1950s, and the latest technologies in the field of stochastic processes and the discovery of Hidden Markov Models have given a new direction for this area.

This paper describes an approach of speech recognition by using the Mel-Scale Frequency Cepstral Coefficients (MFCC) from speech recognition experiments done on OLLO French corpus by different features. Our work consists in finding the most appropriate choice for this task using the Mel-Scale Frequency Cepstral Coefficients (MFCC) extracted from speech signal.

To evaluate this analysis, we built an ASR reference system based on the modeling of phonemes by the HMM (Hidden Markov Models) associated with the GMM models (Gaussian Mixture Model) using the HTK tool. The implementation of this system was made using several experiments in order to choose the best parameters used in two main steps to build an ASR system, acoustic analysis and decoding. The experiments show that the choice of 25 Gaussian components provides a good compromise between recognition accuracy and computation time, and we found also that the best parameters leading to good recognition accuracy are MFCC_E_D_A coefficients with 92.5%.

In this paper the quality and testing of speaker recognition and gender recognition system is completed and analysed.