Use of generalized cepstral distance measure in isolated word recognition

Abstract

This paper proposes a generalized cepstral discance measure based on minimum phase spictral models and shows the performance of the distance measure for isolated word recognition. the generalized cepstral distance measure approximates the L2 norm between two spectra on a fractional power magnitude scale. the measure is equal to the cepstral distance measure for the special case. Calculation of the distance can be performed efficiently from cepstral coefficients or linear predictor coefficients. Isolated word recognition experiments using a vocabulary of 20 highly confusable Japanese city names indicate that the generalized cepstral distance measure gives higher recognition accuracy than conventional distance measures. Three analysis parameter sets, which are FFT cepstral coefficients, linear predictor coefficients, and improved cepstral coefficients, are also compared in terms of their effects on the recognition performance.