Comparative study of various spectrum matching measures on noise robustness

This paper investigates the noise robustness of various LPC and filter bank based spectral matching measures with the aim of developing a reliable speech recognition system for noisy environments. In order to evaluate the recognition performance of the distance measures for noisy speech, speaker dependent isolated word recognition tests in white and colored noise environments were carried out using both clean and noisy reference spectra. The experimental results showed that their recognition performances are very different at large SNR mismatch conditions between test and reference, especially for clean reference, and that the low frequency weighted WLR is the most robust for white noise, and the log likelihood ratio is the worst for highly degraded input speech. Furthermore, pseudo peak weighted likelihood ratios based on filter bank spectra showed very high noise robust characteristics close to the corresponding LPC distance metrics.