Compound wavelets: wavelets for speech recognition

Describes a method for generating a wavelet set that requires a specific time-bandwidth product. This allows independent control of the time resolution and the number of wavelets per time sample of the sampled continuous wavelet transform to accurately parameterise speech for speech recognition. The proposed method increases the bandwidth of a wavelet without significantly affecting the time domain support. Adjacent wavelets are compounded to produce a wavelet with a wider bandwidth. This is performed by summing the wavelet coefficients in the time domain. If the contributing wavelets satisfy the reconstructing admissibility condition, so does the compound wavelet. Speech recognition experiments using the “E-set” (b, c, d, e, g, p, t, v, z) vocabulary were performed. The speech is parameterised with compound wavelets, with increasing compound levels. Recognition performance improves from 66.1% to 71.5% (15% error reduction) by using a compound wavelet composed with two base wavelets. As the compound level increases the recognition performance remains relatively constant