Speech recognition system wherein the consonantal characteristics of input utterances are extracted

Claim 1

A speech-recognition system in which time-series patterns of characteristic quantities are extracted periodically at each frame defined as a predetermined time interval from an input utterance within a voiced interval, i.e. within a time interval from the start point of the utterance until its end point, the similarities are calculated between these time-series patterns and reference patterns prepared in advance, and the similarity is calculated for each category to be recognized, and the category having the largest similarity among all the categories to be recognized is used as the result of the recognition, the speech-recognition system comprising: (a) a spectrum normalizer which performs frequency analysis in a plurality of channels (numbered by their central frequencies) and logarithmic conversion and extracts the frequency spectra, and then calculates normalized spectrum patterns by normalizing the frequency spectra with least square fit lines; (b) a consonantal pattern extractor which makes judgement as to whether each frame has consonantal properties and creates consonantal paterns by processing in sequence the frames within a voiced interval, extracting the consonantal patterns in those frames which are judged to have consonantal properties, and not extracting consonantal patterns in those frames which are judged to lack consonantal properties (i.e., in which the value is set at 0 in all channel components); (c) a local-peak pattern extractor, which creates local peak patterns by processing all the frames within a voiced interval, assigning number 1 to those channel components in which the value of the normalized spectrum pattern is positive and reaches a maximum, and assigning number 0 to all the other channel components; (d) a consonantal similarity degree calculator which calculates the similarity between the consonantal patterns calculated by the extractor in (b) and consonantal reference patterns prepared in advance, and calculates the consonantal similarity for each category to be recognized; (e) a memory unit for the consonantal reference patterns; (f) a local-peak similarity calculator which calculates the similarity between the local-peak patterns calculated by the extractor in (c) and local-peak reference patterns prepared in advance, and calculates the local-peak similarity for each category to be recognized; and (g) an identifier which references both the consonantal similarity and the local-peak similarity and calculates the comprehensive similarity for each category to be recognized, and selecting, among all the categories to be recognized, the category which has the largest comprehensive similarity as the result of recognition.