Dynamic Unit Selection for Very Low Bit Rate Coding at 500 bits/sec

This paper presents a new unit selection process for Very Low Bit Rate speech encoding around 500 bits/sec. The encoding is based on speech recognition and speech synthesis technologies. The aim of this approach is to use at best the speech corpus of the speaker. The proposed solution uses HMM modelling for the recognition of elementary speech units. The HMM are first trained in an unsupervised phase and then are used to build the synthesis unit corpus. The coding process relies on the synthesis unit selection. The speech is decoded by concatenating the selected units through HNM-like decomposition of speech. The new unit selection aims at finding the unit that best match the prosody constraints to models its evolution. It enables the size of the synthesis unit corpus to be independant of the targeted bit rate. A complete quantisation scheme of the overall set of encoded parameters is given.