Corpus based very low bit rate speech coding

This paper presents a new very low bit rate segmental speech coding approach applying speech recognition in the coder and corpus based speech synthesis in the decoder. The system uses a large corpus of speech signals that is searched to find a speech segment similar to the segment to be coded. The elementary acoustical units for recognition and synthesis are determined automatically by an unsupervised training method. This approach is an alternative to using phoneme-derived linguistic units. Very good results are obtained at an average bit rate of 400 bits/second for a corpus of about 1 hour of speech. We present an efficient method for finding the best synthesis unit taking into account the good concatenation of successive segments. The proposed organization of the speech segments in the corpus allows a very efficient search of the best unit.