STRUCTURAL KLD FOR CROSS-VARIETY SPEAKER ADAPTATION IN HMM-BASED SPEECH SYNTHESIS

摘要

While the synthesis of natural sounding, neutral stylernspeech can be achieved using today’s technology, fast adaptationrnof speech synthesis to different contexts and situationsrnstill poses a challenge. In the context of varietyrnmodeling (dialects, sociolects) we have to cope with thernproblem that no standardized orthographic form is availablernand that existing speech resources for these varietiesrnare rare. We present recent approaches in the fieldrnof cross-lingual speaker transformation for HMM-basedrnspeech synthesis and propose a method for transforming anrnarbitrary speaker’s voice from one variety to another one.rnWe apply Kullback-Leibler divergence for data mapping ofrnHMM-states, transfer probability density functions to therndecision tree of the other variety and perform speaker adaptation.rnA method to integrate structural information in thernmapping is also presented and analyzed. Subjective listeningrntests show that the proposed method produces speechrnof significantly higher quality than standard speaker adaptationrntechniques.