Very low bit rate speech coding based on HMMs

Abstract

SUMMARYIn this paper the authors describe very low bit rate speech coding based on HMMs (Hidden Markov Models). In the coder, phoneme recognition is performed using HMMs, then the phoneme index array, state duration length, and pitch information are sent to the decoder. In the decoder, the phoneme HMMs are linked in accordance with the phoneme index array. Then, a mel-cepstral array is generated from the linked HMMs based on a speech parameter generation algorithm using likelihood maximization standards in accordance with the state duration length. Finally, speech is synthesized by exciting an MLSA (Mel Log Spectrum Approximation) filter with the generated mel-cepstral as a coefficient in accordance with the pitch information. The results of subjective evaluation experiments show that performance similar to a vocoder based on 400 bit/s (8 bit/frame u 50 frame/s) vector quantization excluding pitch information can be obtained using the proposed format at 146 bits/s (including a 26% silent interval) excluding pitch information. © 2001 Scripta Technica, Syst Comp Jpn, 32(12): 3846, 2001

Very low bit rate speech coding based on HMM with speaker adaptation

“…Results of phoneme recognition * A recognition error indicates that a speech unit with a higher acoustic likelihood than the correct phoneme was selected. Degradation of recognition accuracy does not necessarily imply that the quality of the coded speech is degraded [16]. Fig.…”

“…The decoder synthesizes speech by connecting the speech units according to the transmitted index and duration. We have also proposed a phonetic vocoder [16] (called an HMM-based phonetic vocoder) in which a speech recognition system based on HMM and the speech synthesis system based on HMM proposed by us [15] are connected. We have shown that the same performance as that of a vocoder based on vector quantization at 400 bits/s (8 bits/frame × 50 frame/s) is obtained at approximately 150 bits/s, excluding pitch data.…”