Hidden Markov Models for artificial voice production and accent modification

Abstract

In this paper, we consider the problem of accent modification between Castilian Spanish and Mexican Spanish. This is an interesting application area for tasks such as the automatic dubbing of pictures and videos with different accents. We initially apply statistical parametric speech synthesis to produce two artificial voices, each with the required accent, using Hidden Markov Models (HMM). This type of speech synthesis technique is capable of learning and reproducing certain essential parameters of the voice in question. We then propose a way to adapt these parameters between the two accents. The prosodic differences in the voices are modeled and transformed directly using this adaptation method. In order to produce the voices initially, we use a speech database that was developed by professional actors from Spain and Mexico. The results obtained from subjective and objective tests are promising, and the method is essentially applicable to accent modification between other Spanish accents.