摘要 |
PURPOSE:To prevent the dropping of the recognition rate of the title system due to the difference in speaker by selecting the sub-set of a standard pattern having a similar inclination in voice vector from the mean value of primary cepstrum coefficients in a voice section. CONSTITUTION:A standard pattern memory section 6 stores standard pattern groups classified into sub-set of different types of feature quantities. A starting/ terminating end detecting section 2 separates a voiced section only and calculates the number of zero-crossing times of a detected voiced section and a feature quantity calculating section 3 calculates the feature quantity of a cepstrum coefficient, etc., from separated voice signals. A pattern selecting section 4 detects a voice section in voice signals on the basis of the number of zero-crossing times found by the section 2 and selects a sub-set having a similar voice spectrum inclination out of a standard pattern set on the basis of the feature quantity of the voice section found by the section 3. In addition, a pattern collating section 5 collates the standard patterns in the selected sub-set with the sequence pattern of the feature quantity calculated by the section. Therefore, voices of different speakers can excellently be recognized. |