Multi stage matrix quantization for very low bit rate speech coding

Abstract

Efficient quantization of linear predictive coding (LPC) filter coefficients play an essential role in very-low-bit-rate speech coding systems. This paper examines a new suboptimal matrix quantization scheme for LPC parameters, called multi-stage matrix quantization (MSMQ), which operates at bit rates between 400 and 800 bit/s. With the new matrix quantization method, using a 22.5 ms LPC analysis frame, spectral distortion about 1 dB is achieved at 800 bit/s. In the proposed coder, line spectral frequency (LSF) parameters of multiple consecutive frames are grouped into a superframe and jointly quantized. The new residual LSF vector quantization scheme gives a bit rate reduction in MSMQ without any additional complexity or storage. The new MSMQ leads into several schemes of various computational complexity/storage characteristics.