COMPARISON OF SUB-MODEL CURTAILING TECHNIQUES TO ACCELERATE VECTOR QUANTIZATION BASED SPEAKER IDENTIFICATION

Automatic speaker identification (ASI) is a remotely operative tacit technique for surveillance and tracking persons through digital telephone networks. Vector quantization (VQ) technique often performs in parity with Gaussian mixture model (GMM) in terms of accuracy and performs better in speed for automated speaker identification (ASI). Real-time speaker identification systems consume most of time comparing d-dimensional feature vectors extracted from a test speech sample with M codewords of codebooks of N registered speakers. Closest codeword search (CCS) is performed for N T times to find the best matching codeword for T number of feature vectors extracted from test speech sample to find the best matching registered speaker. It requires d-dimensional distance computations for M * N * T times. ASI speedup techniques focus on reducing the effect of parameters T, N, M or d. Vantage point tree (VPT) technique tends to reduce M by indexing codeword into a binary tree like structure to speedup CCS. Although best case speedup is expected to be M 2 / log M but best average speedup factor empirically found is reportedly only 1.67 for codebook size M=512. On the other hand partial distortion elimination (PDE) that had been mostly ignored in ASI focuses on reducing d. It has been observed that PDE reduces codebook size M * d by 3 times more than VPT to speedup speaker identification 3 times faster.