Adaptive wavelet extraction for speech recognition.

摘要

Method of extracting features for automatic speech recognition, comprising the steps of: decomposing a signal (30) using a voice set of wavelets; defining a wavelet basis; and applying said wavelet basis to said set (38) of wavelets to generate a plurality of decomposition coefficients representing features extracted from the acoustic speech signal, characterized in that: said step of defining a base (102, 104, 106 , 110, 112, 114, 116, 118) comprises wavelet maximizing discrimination of particular classes of sounds in relation to the decomposed speech signal; said set of wavelets is organized into a tree structure having a root node and a plurality of leaf nodes, wherein a portion of said nodes is selected by computing the discriminative power of a first of the nodes, then computing the discriminative power of child nodes of the nodes of said first and pruning said child nodes if the sum of the discriminative powers of said child nodes is less than the discriminative powers of said first of said nodes; and continuously supplied said acoustic speech signal to the root node of said tree structure.