Toggling and Circular Partial Distortion Elimination Algorithms to Speedup Speaker Identification based on Vector Quantization

Abstract

Vector quantization (VQ) efficiently competes with contemporary speaker identification techniques. However, VQ-based real-time speaker identification systems suffer latency due to distance computation between a large number of feature vectors and code vectors of speakers’ codebooks to find the best match in the database. The identification time depends on dimension and count of extracted feature vectors as well as the number of codebooks. Previous speedup techniques in VQ-based speaker identification decrease test vector count through prequantization and prune out unlikely speakers. However reported speedup factors come with accuracy degradation. This paper proposes techniques to speedup closest code vector search (CCS) based on stationarity of speech. In this paper proximity relationship is substantiated among code vectors extracted through LBG process of codebook generation. Based upon the high correlation of proximate code vectors, circular partial distortion elimination (CPDE) and toggling-CPDE algorithms have been proposed in this paper to speedup CCS. Further speedup is proposed through pruning test feature vector sequence for unlikely codebooks during best match speaker search. Our empirical results show that an average speedup factor up to 5.8 for 630 registered speakers of TIMIT 8kHz corpus and 6.6 for 230 speakers of NIST-1999 database have been achieved through integrating the proposed techniques.