Subspace analysis of spectral features for speaker recognition

A new front-end feature extraction scheme creating so called LDA-projected magnitude spectrum (L-PMS) features is proposed for speaker recognition systems. Mainstream feature extraction schemes usually use filter-bank or linear predictive coding (LPC) in the process of converting high-dimensional speech data into low-dimensional feature vectors, which may lose important discriminative information for speaker recognition tasks. In this work, the new feature extraction scheme takes log of magnitude spectrum of windowed utterance frames. After variance normalization on the spectral features, linear discriminant analysis (LDA) is applied to create discriminatively more powerful features comparing to the conventional mel-frequency cepstral coefficient (MFCC) features. The new feature was evaluated on the TIMIT and NTIMIT corpora, using vector quantization (VQ) speaker model. The Experiments on all 630 subjects in TIMIT and NTIMIT corpora show that the proposed L-PMS features substantially outperform the conventional MFCC features in the sense of identification rate.