Robust Speaker Verification Based on Multi Stage Vector Quantization of MFCC Parameters on Narrow Bandwidth Channels

Abstract

This paper presents a very low bit rate and robust client-server-based speaker verification system using MFCC parameters. Two aspects are proposed and assessed including very low bit rate transmission of test utterance feature vectors from client to server, and robust speaker verification in situations where training and test environment noise conditions including noise types and SNRs are different and unknown for speaker verification system. Very low bit rate transmission of feature vectors are achieved using multi stage vector quantization technique (MSVQ). This technique is used for quantization of MFCC feature vectors obtained from speaker's utterance in client side. This leads to significant bits per frame (bpf) reduction from 416 bpf for transmission of 13 dimensional MFCC feature vectors to 36 bpf i.e. 3600 bps. Robust speaker verification is achieved when instead of training only a speaker model using clean data, several speaker models are trained using a limited number of noises in different SNRs. This leads to very good performances even for conditions where test environment noise types and SNRs are different from those of training phase. The results of conducted experiments approve the effectiveness of the proposed methods.