Using resynthesized speech in /r/–/l/ production and perception training

Abstract

To find effective speech stimuli for second-language (L2) training, we generated resynthesized speech using the STRAIGHT algorithm developed by Kawahara (1997). First, Japanese speakers were trained to produce English /r/ and /l/ in a reproduction task. Training proceeded by first using model words which were artificially lengthened by STRAIGHT, and then gradually introducing words at a natural speed. Spectrograms of the model speech and the learner’s speech were given as feedback. Subjects improved significantly from pretest to post-test in their production ability as evaluated by human judges. Second, resynthesized speech was used in perception training. Previous perceptual training studies using synthetic speech have not always been successful. However, training based on resynthesized speech from STRAIGHT led to the same amount of improvement as training based on the original natural tokens, suggesting the potential for using STRAIGHT in L2 perception training. We then manipulated certain acoustic parameters of the stimuli, namely, lengthening and/or enhancement of the third formant. The training effect using these stimuli did not differ significantly from the effect found with the unmodified synthetic stimuli. These results suggest that resynthesized speech from STRAIGHT is suitable for L2 training, but further effort is necessary to find more effective stimulus manipulations.