摘要 |
PURPOSE:To perform voice recognition based upon a more accurate voice pattern by reducing the influence of a noise due to the circuit system of the device and a dark noise in circumferential environment as much as possible. CONSTITUTION:A specific voice 'stop' is inputted to a microphone 3 before the voice recognition. A frequency analysis part 6 finds the power spectrum of a noise from the voiceless section between 'to' and 'p', frequency by frequency, and this is stored in the calibration register 13 of a feature quantity extraction part 11 to update a correction quantity. Then the value of the power spectrum of a noise part stored in the calibration register 13 is subtracted from the value of the power spectrum analyzed by the frequency analysis part 6 to correct the voice to be recognized. A feature quantity extracting circuit 14 extracts a feature quantity required for the voice recognition from respective power spectra after the correction. |