FFTに基づくBPFを用いた韓国語数字認識

概要

論文の詳細を見る
This paper describes the recognition system of Korean digits spoken in isolation. Especially investigated are the bandwidth and the center frequency allocation of filters for extracting the spectral envelope from the power spectrum obtained by fast Fourier transform(FFT). This experiment is carried out as a preliminary investigation of the recognition of Korean speech. The filters are realized by multiplying the rectangular frequency windows by the spectra obtained by FFT. Regarding the bandwidth of band pass filters, three types of filters are compared. The first one is of the constant frequency bandwidth in the linear frequency scale. The second one is of the constant bandwidth in the linear-logarithmic frequency scale(linear in the range lower than about 1, 000 Hz and logarithmic in the higher frequency range). The third one is of the constant bandwidth in the logarithmic frequency scale. Experiments are carried out on both speaker dependent and speaker independent recognition of 10 Korean digits uttered by five males and five females(1, 000 test samples). On the speaker dependent mode a 99. 3% correct recognition rate is obtained using 9 channel filters whose bandwidths are constant in the logarithmic frequency scale. The same rate is obtained using 12 channel filters whose bandwidths are constant in the linear-logarithmic frequency scale.
社団法人日本音響学会の論文