| 摘要 |
<p>PURPOSE:To reduce speaker's variation absorption or the volume of calculation and to attain the efficiency of the system by matching a feature pattern formed from an input voice signal through an acoustic analyzing part and a pattern conversion part with a reference pattern obtained from a storage part. CONSTITUTION:A voice signal inputted from a microphone 1 is analyzed by the acoustic analyzing part 2 and the time sequence of acoustic parameters is outputted in each unit time. The time sequence is inputted to the pattern conversion part 3 and the time sequence of articulation vectors consisting of elements such as narrow degree, an articulation point, the degree of nasalization, the existence of vocal cord vibration and degree of round is formed by a procedure using a neural net or the like. The articulation vector time sequence is found out in each voice sample based upon the feature parameters obtained by acoustic analyzing the voice sample. A representative articulation vector time sequence is led out by statistic processing and the formed reference pattern is stored in a reference pattern storing part 6. Then, DP matching between the articulation vector time sequences of both the patterns is executed by an identification part 5. When there are many vocabularies to be recognized, the reference patterns to be matched in the ID part 5 are preparatorily selected by a preparatory selection part 4 to converge the pattern.</p> |
