Speaker Recognition by Gaussian Filter Based Feature Extraction and Proposed Fuzzy Vector Quantization Modelling Technique

摘要

Automatic speaker-recognition systems (ASR) have developed as a vital method for recognition of a person in many applications like e-commerce as well as in common interactions, law enforcement, and forensics. The performance of automatic speaker recognition system depends on the duration of the voice of test and train samples. In this paper, I address the issues in the design of codebook for outlier effect using the fuzzy vector quantization based algorithm for tasks such as speaker recognition for short voice samples. Particularly, I address the issues in partitioning the voice data for matching sequences of speaker specific feature vectors extracted from the voice signal data of utterances of a speaker. The design of codebook based on fuzzy vector quantization (FVQ) and fuzzy c-means (FCM) vector quantization has been proposed previously for matching the test and train voice signal characterized as sets of speaker-specific features vectors for tasks such as distortion measure and speaker recognition. Speaker-specific features Mel Frequency Cepstral Coefficients (MFCC) extracted by the traditional triangular filter as well as Gaussian and Tukey shapes filter. This research paper compares experimental results of three different modelling techniques namely, Fuzzy c-means, Fuzzy Vector Quantization2 (FVQ2) and proposed fuzzy vector quantization (FVQ). ASR efficiency of FVQ shows significant improvement in performance compared to FCM and FVQ2. The efficiency of proposed ASR system is 98.8% for 2 seconds of training voice data for a set of 100 speakers taken from the Texas Instruments and Massachusetts Institute of Technology (TIMIT) database.