Minimization of transition noise and HNM synthesis in very low bit rate speech coding

speech coding, speech recognition, speech synthesis, feature extraction

Minimization of transition noise and HNM synthesis in very low bit rate speech coding

The aim of our effort is to reach higher quality of resulting speech coded by very low bit rate (VLBR) segmental coder. In already existing VLBR coder, we want to improve the determination of acoustical units. Furthermore, better analysis-synthesis technique for the synthesis part (Harmonic-Noise Model) instead of LPCC is going to be used. The VLBR coder consists of a recognition system followed by a speech synthesizer. The recognizer identifies recognition acoustic units (RU). On the other hand, the synthesizer concatenates synthesis acoustic units (SU). However, the two kinds of acoustic unit can be identical or different and then can be modeled in different ways such Hidden Markov Model for the RU and Harmonic-Noise model for the SU. Both kinds of units are obtained automatically from a training database of raw speech that does not contain any transcription. In the original version of the coder, the quality of the synthetic speech was not sufficient for these two main reasons: the SU units were too short and difficult to concatenate and the synthesis was done using basic LPCC analysis-synthesis. In order to remove first drawback, three methods of re-segmentation were used. Afterwards, the basic LPCC analysis-synthesis was replaced by HNM.