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.
Muhammad Afzal, Mohammad A. Maud, Ali Hammad Akbar. (2012) Toggling and Circular Partial Distortion Elimination Algorithms to Speedup Speaker Identification based on Vector Quantization, Pakistan Journal of Engineering and Applied Sciences, Volume 11, Issue 1 .
-
Views
2035 -
Downloads
197
Next Article
Article Details
Volume
Issue
Type
Language
