No Detail Found!

Efficient vq-based speaker identification through elimination of anomalous vectors during mfcc feature extraction.

Author(s)
M. Afzal Department of Computer Science and Engineering, University of Engineering & Technology, Lahore, Pakistan
T. Ahmad Department of Computer Science and Engineering, University of Engineering & Technology, Lahore, Pakistan
M. F. Hayat Department of Computer Science and Engineering, University of Engineering & Technology, Lahore, Pakistan
H. M. S. Asif Department of Computer Science and Engineering, University of Engineering & Technology, Lahore, Pakistan
K. H. Asif Department of Computer Science and Engineering, University of Engineering & Technology, Lahore, Pakistan
Abstract
Floating Point (FP) numbers are ubiquitous to manipulate real number in digital computations. Mel-Frequency Cepstral Coefficients (MFCC) feature vectors of real values are extracted from speech to represent a large number of phonetic variations pertaining to speaker and spoken text. The MFCC vectors require log, sine, cosine and etc for their computation. However, multilevel rounding that occurs in FP arithmetic operation requires extra care for computing MFCC vectors. Computing log of FP zero gives FP -¥, which is carried along till last step in extracting MFCC vectors. We propose elimination of anomalous vectors in MFCC features vectors extraction process. Vector Quantization (VQ) compresses MFCC vectors into a small set of mean vectors called codebook to efficiently represent phonetic classes thus making them more feasible to process speech for speaker verification and identification etc. Average taking in clustering algorithms of FP numbers containing FP ¥ -¥ computes to FP -¥ that give false codebook. Removal of vectors containing FP element is proposed for speaker identification and other VQ-based classification systems in general for higher accuracy. Accuracy was increased by 2%, 3% and 2% in our VQ based speaker identification experiments on TIMIT 16 kHz, down sampled TIMIT 8 kHz and CSLU speaker recognition data respectively.
Publication Details
Page(s) 36-38
DOI DOI not available
Published Journal: Pakistan Journal of Science, Volume: 66, Issue: 1, Year: 2014

Request the full text from PASTIC

0

Citations

0

Downloads

35

Views