Speech Recognition Using Multiple Features and Multiple Recognizers

摘要

The purpose of this thesis is to demonstrate the feasibility of using multiplefeatures and multiple recognizers to perform isolated word recognition. This is accomplished by performing multiple independent recognition tests and fusing the results together to get a single recognition result. The speech data is recorded and each word is extracted into a separate file. Eight features are calculated for each word. The features are calculated on 512 sample time slices and produce 16 component vector output. The three recognizers use the eight features to produce a total of 24 error distance lists. These lists are then fused together by adding the error values corresponding to each word. The word with the smallest fused error value is declared the recognition winner. Talker dependent and independent tests were run on a word set of zero through nine and A through Z. The talker dependent tests achieved accuracies between 87% and 100% depending on the talker. The talker independent tests achieved accuracies between 81% and 97%.