Effect of temporal fluctuation in speech on perception of humanness of synthesized speech

Abstract

The sound quality of speech synthesized using modern speech synthesis systems is expected to be approximated to human speech. We investigated the effect of temporal fluctuation of speech on the perception of humanness. Speech stimuli used in the evaluation were generated by voice morphing. The morphing source (morphing rate of 0%) and target (morphing rate of 100%) were speech without temporal fluctuation and original speech, respectively. There were three kinds of morphing factors: fundamental frequency (F0), spectral envelope (SP), and both (F0 + SP). We used MUSHRA as the evaluation method involving two speakers (one male and one female), two phonemes (/a/ and /i/), and two F0s (high and low). Nine morphing rates (every 25% from -100 to 100%) were used, and ten subjects with normal hearing participated in the evaluation. The results show that the stimuli with a morphing rate of 0% scored the lowest humanness for all factors. The stimuli with morphing rates of -100% scored significantly lower than those with a morphing rate of 100% in SP and F0 + SP. The most dominant factor was F0 + SP, and the effect of F0 was the smallest of all factors. [Work supported by JSPS KAKENHI Grant Numbers 15H02726, 16H05899, 16H01734.]