Current Status of the IBM Trainable Speech Synthesis System

Abstract

This paper describes the current status of the IBM Trainable Speech Synthesis System. The system is a state-of-the-art, trainable, unit-selection based concatenative speech synthesiser. The system uses hidden Markov models (HMMs) to provide a phonetic transcription and HMM state alignment of a database of single-speaker continuous-speech training data. The runtime synthesiser uses the HMM state sized segments that result as its basic synthesis units. It determines which segments to concatenate to produce a target sentence using decision trees built from the training data and a dynamic programming search to optimise a perceptually motivated cost function. The synthesiser can operate both in general domain Text-to-Speech mode, and in Phrase Splicing mode to provide higher quality synthesis in limited domains. Systems have been built in at least 10 different languages and over 70 voices.