| 摘要 |
PURPOSE:To provide the speech recognizing method which can improve the performance of the automatic speech recognition by a machine by calculating a spectrum time sequence more approximated to human hearing rather than the conventional method by using the principle of a time frequency filter having pass characteristics depending on continuous time. CONSTITUTION:The speech inputted to a microphone 1 is converted to a digital signal by an A/D converter 3, a self-correlation coefficient is obtained by a self-correlation analysis part 4, and a linear prediction coefficient is obtained by a linear prediction analysis part 5. Further, a cepstrum coefficient is calculated by a cepstrum analysis part 6, and a dynamic cepstrum time sequence is obtained by performing time frequency masking filtering to cepstrum time sequence at a dynamic cepstrum generation part 7. A switch SW 1 is changed over, the vector quantization of the dynamic cepstrum is performed to store in a code book storage part 9, the speech is expressed by a vector code sequence in a vector quantization part 10, and HMM obtained by HMM learning is stored in an HMM storage part 12 and recognized by an HMM recognition part 13. |
