An Analysis of Wavefunction-Derived Speech Recognition Features

Abstract

Based upon a wavefunction representation of wide-band filtered speech (five fixed-frequency bands), recognition features can be computed pitch synchronously and then later combined at the phoneme level. The pitch synchronous wavefunction frequency data at times acts as a formant tracker and at other times does not. The phoneme features from wavefunctions cluster much the same as equivalent formant features but not necessarily at the same locations in feature space. For certain types of sounds (e.g., /aɔ/), the detailed description provided by the wavefunction parameters can be useful in distinguishing one from the other. These and other relationships between formant and wavefunction based recognition features will be discussed.