Accelerating Vector Quantization Based Speaker Identification

Matching of feature vectors extracted from speech sample of an unknown speaker, with models of registered speakers is the most time consuming component of real-time speaker identification systems. Time controlling parameters are size and count of extracted test feature vectors as well as size, complexity and count of models of registered speakers. We studied vector quantization (VQ) for accelerating the bottlenecking component of speaker identification which is less investigated than Gaussian mixture model (GMM). Already reported acceleration techniques in VQ approach reduce test feature vector count by pre-quantization and reduce candidate registered speakers by pruning unlikely ones, thereby, introducing risk of accuracy degradation. The speedup technique used in this paper partially prunes VQ codebook mean vectors using partial distortion elimination (PDE). Acceleration factor of up to 3.29 on 630 registered speakers of TIMIT 8kHz speech data and 4 on 91 registered speakers of CSLU speech data is achieved respectively. [Muhammad Afzal, Shaiq A. Haq. Accelerating Vector Quantization Based Speaker Identification, Journal of American Science 2010;6(11):1046-1050]. (ISSN: 1545-1003). http://www.americanscience.org.