Development of a real-time speaker recognition system using TMS320C31

Abstract

Speaker recognition systems based on Malay language have been developed in the personal computer environment. This thesis outlines a hardware implementation of a real-time speaker recognition using Malay language. Various speaker recognition classifiers have been investigated in term of feasibility in a stand-alone hardware platform. Computational and memory requirement are given consideration, along with processing optimizations. A speaker recognition board is implemented based on a TMS32C31 digital signal processor (DSP). The speaker recognition techniques used are the Linear Predictive Coding (LPC) Cepstral analysis for feature extraction, Vector Quantization (VQ) for feature compression and the Dynamic Time Warping (DTW) for speaker feature matching. This system is trained and tested using a population of ten users, with additional testing using ten impostors. The average entry success of a true user is 93.4%. The speaker recognition board is successfully tested as a speaker recognition door access system, with true access success rate of 88.7%. The speaker recognition system shows good performance, as well as being operational in real-time.