Novel Detection Algorithm of Speech Activity and the impact of Speech Codecs on Remote Speaker Recognition System

Abstract

In this paper, we studied the effects of voice codecs on remote speaker recognition system, considering three types of speech codec: PCM, DPCM and ADPCM conforming to International Telecommunications Union - Telecoms (ITU-T) recommendation used in telephony and VoIP (Voice over Internet Protocol). To improve the performance of speaker recognition in a noisy environment, we propose a new speech activity detection algorithm (SAD) using ''Adaptive Threshold'', which can be simulated with speech wave files of TIMIT (Texas Instruments Massachusetts Institute of Technology) database that allows recognition system to be done under almost ideal conditions. Moreover, the speaker recognition system is based on Vector Quantization as speaker modeling technique and Mel Frequency Cepstral Coefficient (MFCC) as feature extraction technique. Where, the feature extraction proceed after (for testing phase) and before (for training phase) the speech is sending over communication channel. Therefore, the digital channels can introduce several types of degradation. To overcome the channel degradation, a convolutional code is used as error-control coding with AWGN channel. Finely, In our simulation with Matlab we have used 30 speakers of different regions (10 male and 20female), the best overall performance of speech codecs was observed for the PCM code in terms of recognition rate accuracy and runtime