| 摘要 |
PROBLEM TO BE SOLVED: To provide a device and a method for improving the recognition performance of narrow-band voice recognition almost to that of wide-band voice recognition. SOLUTION: A feature extraction part 10 is equipped with a power spectrum calculation part 11 which calculates the short-time power spectrum of an inputted narrow-band voice, band conversion part 13, and a malcepstrum calculation part 12 which calculates a malcepstrum feature variable by using a band- widened power spectrum and a band conversion part 13 has a characteristic vector storage part (33 in Fig. 3) which stores the characteristic vector of the power spectrum of the wide-band voice, an expansion coefficient calculation part (31 in Fig. 3) which calculates an expansion coefficient obtained by expanding the power spectrum of the narrow-band voice calculated by the power spectrum calculation part with the characteristic vector, and a band expansion part (31 in Fig. 3) which calculates the power spectrum of an absent band part by using the expansion coefficient and combines it with the narrow- band spectrum to calculate a dummy wide-band voice power spectrum. |
