Evolutionary weight tuning based on diphone pairsfor unit selection speech synthesis

Abstract

Unit selection text-to-speech (TTS) conversion is an ongoingresearch for the speech synthesis community. This paper isfocused on tuning the weights involved in the target and con-catenation cost metrics. We propose a method for automati-cally adjusting these weights simultaneously by means of di-phone and triphone pairs. This method is based on techniquesprovided by the evolutionary computation community, takingadvantage of their robustness in noisy domains. The experi-ments and their analyses demonstrate its good performance inthis problem, thus, overcoming some constraints assumed byprevious works and leading to a new interesting framework forfurther investigations.