Sixth International Conference on Spoken Language Processing
(ICSLP 2000)

Beijing, China
October 16-20, 2000

Efficient Segment Quantization of LSP Parameters for Very Low Bit Speech Coding

Minoru Kohata (1), Ikuya Mitsuya (2), Motoyuki Suzuki (2), Shozo Makino (2)

(1) Chiba Institute of Technology, Narashino, Japan
(2) Tohoku University, Sendai, Japan

This paper presents a new segment quantization (SQ) method for LPC coefficients, which is based upon a new segmentation scheme. In order to design an efficient segment coder, a segmentation method using the self-similarity of LSP coefficients between several frames is employed. The segmentation is carried out by matching the input LSP frames with each segment in the codebook, and the codebook is trained at the same time with a method similar to the Lempel-Ziv coding method, which is one of the universal coding methods for discrete symbols. In the training process, the segment codebook is grown from null to the desired size. Two types of segment quantizer are designed based upon the proposed segmentation method, and both methods can operate at low rates of below 7 bit/frame, with low complexity and low cepstral distortion of less than 2.4 dB.


Full Paper

Bibliographic reference.  Kohata, Minoru / Mitsuya, Ikuya / Suzuki, Motoyuki / Makino, Shozo (2000): "Efficient segment quantization of LSP parameters for very low bit speech coding", In ICSLP-2000, vol.3, 826-829.