Tree-based context clustering using speech recognition features for acoustic model training of speech synthesis

Tree based context clustering processes reduce the sizes of acoustic models of Hidden Markov Model (HMM) speech synthesis systems as well as eliminate problems arising from unseen sound units. Representations of speech units in speech synthesis systems are often LPC or MCEP features whose characteristics promote speech reconstruction rather than discrimination among different sound units. In this paper, MFCC features, successfully utilized in speech recognition, were selected as features for generating context clustering trees applied to LPC/MCEP-based speech synthesis. On average, the collective size of acoustic models was 29% smaller than ones of typical cases while spectral features generated from a speech synthesis system using each type of clustering trees did not significantly deviate from features extracted from actual spoken utterances. Applying MFCC-based clustering tree did not significantly affect the resulting pitch and duration models of the system. We concluded that MFCC-based clustering tree can reduce the overall size of acoustic models while synthetic sound quality is maintained.