Multi-mode matrix quantizer for low bit rate LSF quantization

In this paper, we introduce a novel method for quantization of line spectral frequencies (LSF) converted from mth order linear prediction coefficients. In the proposed method, the interframe correlation of LSFs is exploited using matrix quantization where N consecutive frames are quantized as one m-by-N matrix. The voicing-based multi-mode operation reduces the bit rate by taking advantage of the properties of the speech signal. That is, certain parts of a signal, such as unvoiced segments, can be quantized with smaller codebooks. With this method, very low variable bit rate LSF quantization is obtained. The proposed method is suitable especially for very low bit rate speech coders in which short time delay is tolerable, and high but not necessarily transparent quality is sufficient.