5^{th} International Conference on Spoken Language ProcessingSydney, Australia |
A method of inter-frame transform coding of Line Spectrum Frequencies (LSF) using the Discrete Wavelet Transform is presented in this paper. Each component of the LSFs (or of their linear transform) is treated separately and is decomposed into a set of subband signals using the nonuniform filter bank. Subband signals are quantized and coded independently. By the appropriate choice of the mother Wavelet, subband signal with the lowest rate comprises most of the LSF waveform energy. Filter bank effectively decorrelates the input signal, enabling more efficient quantization of the subband signals. A suitable weighted Euclidean distance measure in the Wavelet domain is proposed, defining optimal static or dynamic bit allocation of the subband signals. It is shown that the average bit rate for coding of the DCT transformed LSFs can be reduced by 0.9 bits per vector component by using a very simple Wavelet. The total delay due to the inter-frame coding is only 90ms that is acceptable even for a medium bit rate speech coders.
Bibliographic reference. Petrinovic, Davor (1998): "Transform coding of LSF parameters using wavelets", In ICSLP-1998, paper 1114.