ADAPTIVE VOICE RECOGNITION DEVICE

摘要

PURPOSE:To attain the high-performance adaption of a speech pattern wherein the environment of preceding and following phonemes having no influence upon an error in quantization can be considered by converting a standard pattern generated by the voice of a standard speaker into the standard pattern of a specific speaker by using a neutral network and recognizing the speech of the specific speaker with the converted standard pattern. CONSTITUTION:A time-base matching part 13 finds the feature vector b(j(i)) of a learning pattern B corresponding to the feature vector a(i) of the standard pattern A. A learning control part 14 outputs a frame number (i) to a read-out control part 15 first when the feature vector b(j(i)) is inputted. A standard pattern a(k) nearby the frame number is read out of a standard pattern memory part 11 with a control signal k(i-e<=k<=i+e) from the read-out control part 15. When the learning of a neutral network part 10 is finished, all standard patterns A in the standard pattern memory part 11 are converted by the neutral network part 10 and held in an adaptive standard pattern memory part 16.