ミスマッチ陰性電位を使用した合成音声の評価と考察(一般講演A,脳計測)

抄録

The purpose of our study is to evaluate the quality of the synthesized speech by measuring the mismatch negativity (MMN) of the auditory event-related potentials (ERPs). The plosive consonant-vowel (CV) pair/tsu/was sinusoidally interpolated by using two kinds ofextrema. One was exact acquisition in the filtered speech with a passing bandwidth of one octave and a sampling frequency of 1 MHz, and another was approximation in the filtered wav-formatted data. Subjects were presented with a repetitive stimulus of a conventional recording /tsu/, randomly replaced at a 14.3 % probability by a deviant stimulus. As compared with the MMN amplitude when presenting a pure tone for the deviant stimulus, the modulus of MMN is significantly reduced by 4.97μV when using the exact synthesis and by 3.96μV when using the approximation. These results have interpreted as the evidence, in which the synthesized speeches were phonetically equivalent to the conventional recording.