A two-dimensional cepstrum approach for the recognition of mandarin syllable initials

摘要

A two-dimensional cepstrum (TDC) approach for the recognition of Mandarin syllable initials is presented. The TDC has the characteristics that it explicitly incorporates the correlations of neighboring frames and well represents the spectral information of the utterance. In this approach, the average magnitude difference function (AMDF) is first used to determine a time instant which will be treated as a boundary of the initial in a Mandarin syllable. The waveform starting from the beginning of the syllable to the boundary is treated as a recognition unit for the initials. The TDC matrix of the recognition unit is then calculated. It is shown that only a small part of the TDC matrix is necessary for the representation of the syllable initial. In the recognition phase, the Bayes rule is adopted. The experimental results demonstrate the effectiveness of the proposed method.