Use of instantaneous and transitional spectral information in speech recognizers.

摘要

Any of several prior art word recognizers are improved by changing the speech analysis portion thereof to include a more comprehensive spectral representation which can be characterized as two-dimensional. The first dimension is the "freeze-frame" or instantaneous sample of the spectral characteristics as derived from a typical 45 millisecond sample of the speech to be recognized. The second dimension, in a sense orthogonal to the first, spans several such time frames or samples and yields what amounts to a time derivative of the spectral properties obtained in the current time frame. In the testing or processing of the input speech in comparison to the reference patterns, essentially equal weight or effectiveness is attributed to both parts of the two dimensional reference spectral pattern and to both parts of the input speech spectral pattern. That is, equal importance is attributed to the essentially instantaneous information and the time-derivative information spanning a plurality of neighboring time frames including the frame of interest. Also disclosed are the use of cepstral information and time-derivative cepstral information derived from linear prediction coefficient methods and the use specifically in connected-word recognizers using level-building concepts, or in those using two-stage processing.