Improvement of synthetic speech quality using a new spectral smoothing technique

초록

Most of speech synthesis systems are basically accomplished by concatenating acoustic units selected from large database. However, this method has a limitation that discontinuities often occur between the units because it is impossible to collect all required variations of the units in the finite database. To solve this problem, this paper proposes a new spectral smoothing technique which is reflecting not only formant trajectory and distribution characteristic of spectrum but also human's acoustic characteristics. The proposed technique decides the quantity and the extent of smoothing by considering human's acoustic characteristics at the concatenation part of unit phonemes, and then performs the spectral smoothing using weights calculated along a time axis at the border of two diphones. Through this approach, we can reduce the discontinuity and minimize the distortion which can be occurred by spectrum smoothing. For the performance evaluation, we tested on five hundred diphones which is extracted from twenty sentences in a voice database and self recorded samples in individual.