Rechercher des projets européens

Rapid Cross-Lingual Speaker Adaptation for Statistical Text-to-Speech Systems (CLSASTS)
Date du début: 1 févr. 2011, Date de fin: 31 janv. 2015 PROJET  TERMINÉ 

"Unit selection has been the dominant approach to text-to-speech synthesis (TTS) in the last decade. Recently, statistical TTS (STS) is proposed where statistical models are used for speech synthesis. The high quality and intelligibility speech it generates, the flexibility it offers in voice/speaker/emotion conversion, and its small memory requirements make STS systems a strong candidate to be the dominant TTS technology in the next decade. One of the most exciting research directions in the STS field is speaker adaptation where the goal is to adapt the voice characteristics to a target speaker. Maximum a posteriori and maximum likelihood linear regression methods are two of the common approaches for adaptation. There is also a recent and growing interest in cross-lingual speaker adaptation using the STS approach where the goal is to use speaker adaptation techniques to generate speech with a speaker’s voice characteristics in a target language that the speaker does not speak. Globalization and the need to communicate in multiple languages in social, economic, and political interactions increase the importance of the problem. In this proposal, a novel rapid adaptation approach is proposed for the cross-lingual speaker adaptation problem using the eigenvoice adaptation technique with the STS systems. In the proposed approach, two sets of eigenvoices are produced, one for the target language and one for the source language. Then, a regression function is generated between the source and target eigenvoice weights. Given a speaker, eigenvoice weights for the source eigenvoices are computed using a novel, perceptually-motivated objective function, and then the regression function is used to estimate the target eigenvoice weights which are then used to synthesize speech in the target language. The proposed system is expected to be the first high-performance cross-lingual speaker adaptation method for STS that can work with 5-10 seconds of adaptation data."

Coordinateur

Details