Characterising speech data for speaker veri cation

Abstract

Speech samples of xed one word were analysed for multiple utterances by the single speaker as well as by di erent speakers. Data were collected from six speakers with 90 speech samples of each. From which, 70 samples were used for training and 20 samples were used for testing. Some of the basic features analysed were, fundamental frequency, amplitude values at highest peaks of FFT, quiet regions and noisy regions for collected data. Some of the features appear to be useful for veri cation of speci c speakers. Mel Frequency Cepstrum Coe cients (MFCC) and Vector Quantization (VQ) derived from the MFCC were applied for the purpose of speaker veri cation using Support Vector Machine (SVM) algorithm. The computational results using di erent parameters like di erent sizes of training samples, di erent codebook size, di erent value of cepstral coe cients, frequency limits and di erent sizes of windows are reported. Using MFCC and VQ, for the training sample of size 70, it was possible to correctly classify all 20 test samples