HMM-based Speech Synthesis: a brief overview

During the last decade, speech synthesis technologies have evolved from concatenative unit-selection techniques to statistical parametric ones. Nowadays, the most relevant statistical technique is HMM-based speech synthesis, where hidden Markov models are used to model phonetic units and an estimation method, such as maximum likelihood, is used as criterion for acoustic model training and parameter generation. The main advantage of this technique is its enormous flexibility for modifying the speech characteristics by applying adaptation techniques. In addition to the speaker identity modification, this flexibility allows obtaining synthetic voices with different speaking styles, emotions, etc. without a costly process of recording new speakers or speaking styles. Other benefits are a low memory footprint and a more consistent synthetic speech quality. New applications like personalized speech-to-speech translation or voice reconstruction for speech impaired people are now possible thanks to the emergence of statistical parametric synthesis. This presentation intends to give a brief overview of the HMM-based speech synthesis and also to show some work carried out in the field of speaker adaptation.