摘要 |
PURPOSE: To make a device for a standard speaker rapidly learn by a later speaker by interpolating the basic probability formed in accordance with the few sample data of the later speaker and the smoothing probability by a confusion matrix to obtain a final parameter. CONSTITUTION: This apparatus carries out such speech recognition that each of the time intervals to be continued corresponds to one label output of label output alphabets (set) and that words correspond to the sequence of a Markov model phone machine. The apparatus has an acoustic processor 102, a stack decoder 104, an acoustic match processor 106 and a language model processor 108. The apparatus successively forms the labels at every time interval to be continued according to speech input, outputs label strings and calculates the basic label output probability from the labels formed in accordance with the utterance of the later speaker. Further, the apparatus forms the final label output probability after forming the smoothing label output probability, based on the similarity between the later speaker and the standard speaker. |