Because of their high correlation and reduced sensitivity to quantization errors, the line spectrum pair (LSP) frequency parameters have been used recently for efficient quantization of LPC information for speech coding. In the present paper, the LSP representation is used for speech recognition and a new perception‐based LSP distance measure is proposed. This distance measure exploits the following two properties of the speech perception process [D. H. Klatt, Proc. ICASSP, 1278–1281 (1982)]: (1) The formant frequencies are the most important parameters for speech perception: and (2) the formant bandwidths and spectral tilt contribute very little to speech perception. The present distance measure uses these two properties in the form of weights in a weighted Euclidean distance measure. In order to derive these weights, the LPC power spectrum P(f) is computed for each speech frame and the weight for a given LSP frequency ft is taken to be proportional to [P(ft)]c. The perception‐based LSP distance measure is studied here on a speaker‐dependent speech recognition task. The optimum value of the exponent c is found here to be 0.15. The perception‐based LSP distance measure results in a recognition score of 95.7%, while the recognition accuracy is found to be 89.9% by using the conventional Euclidean distance measure on LSP parameters and 94.1% by using the liftered cepstral distance measure [B. H. Juang et al., IEEE Trans. Acoust. Speech Signal Process. ASSP‐35, 947–954 (1987)].

This content is only available via PDF.