Distance measure for speech recognition based on the smoothed group delay spectrum

Abstract

We present a novel spectral distance measure based on the smoothed LPC group delay spectrum which gives a stable recognition performance under variable frequency transfer characteristics and additive noise. The weight of the n-th cepstral coefficients in our measure is given by W_{n} = n^{s}. \exp(-n^{2}/2\tau^{2}) which can be adjusted by selecting proper values of s and τ. In order to optimize the parameters of this distance measure, extensive experiments are carried out in a speaker-dependent isolated word recognition system using a standard dynamic time warping technique. The input speech data used here is a set of phonetically very similar 68 Japanese city name pairs spoken by male speakers. The experimental results show that our distance measure gives a robust recognition rate in spite of the variation in frequency characteristics and signal to noise ratio(SNR). In noisy situations of segmental SNR 20 dB, the recognition rate was more than 13% higher than that obtained by using the standard Euclidean cepstral distance measure. Finally, it is shown that the optimum value of s is approximately 1, and the optimum range of τΔT is about 1 ms.