電話音声を対象とした話者照合

概要

論文の詳細を見る
Speaker verification is the only method feasible for remote telephone handset user identification without resorting to additional input devices. Studies are made of speaker verification using the microphone input voice as the reference frame for speaker verification of an actual telephone voice. Various parameter sets as autocorrelation coefficients, linear prediction coefficients, PARCOR parameters, and log area ratio are evaluated for several analysis conditions in order to obtain the best parameter set and conditions. A score as high as 99. 8% is obtained with PARCOR parameters under the best analysis condition (see Table 4). Applying the above best parameter set and conditions to actual telephone voice verification directly, poor results of only 65% are obtained. Then a new idea involving average-self-inverse filtering has been proposed which normalizes the effect of linear distortion of actual telephone voice. The combined use of average-self inverse filtering and selected-band-analysis improves verification performance beyond that obtained with PARCOR parameters extracted directly from an actual telephone voice. Further improvement is possible by the additional use of pitch frequency for normalization of nonlinear distortions so that a better than 94% verification score is obtained.
社団法人日本音響学会の論文
1979-02-01