A new approach of LPC analysis based on the normalization of vocal-tract length

An approach to linear prediction coefficient (LPC) analysis based on the normalization of vocal-tract length is presented. The approach is of significance for speech recognition of arbitrary speakers. In this approach, the ratio of two vocal-tract lengths corresponding to a new speaker and a reference one is first estimated from the training speech data of several typical vowels. The LPC parameters normalized on this ratio can then be calculated for any speech data. Compared with previous methods of speech parameter normalization, this approach does not need to estimate formant frequencies and is simple and reliable in theory. Limited experiments on the recognition of nine Chinese vowels for four speakers to indicate that this new approach can achieve 5% to 20% improvements of correct recognition rate