Third International Conference on Spoken Language Processing (ICSLP 94)
In this paper, a two-stage approach based on KLT transform and 2D prediction is proposed for encoding the LSP parameters. At the first stage, KLT transform is applied to the even part of LSP parameters which are mean-removed. Then DPCM coding is performed on the transformed parameters. At the second stage, the even part of LSP parameters is reconstructed, and a third-order two-dimensional prediction is used to estimate the odd part of LSP parameters. This approach considers both interframe and intraframe correlations simultaneously. The simulation on the database provided by different speakers shows that 1 dB2 difference limen of spectral distortion can be achieved at 19 bits per frame by using the partitioned vector quantization (PVQ) to the residuals of LSP parameters. The residuals of LSP parameters are partitioned into even part and odd part according to the processing of stage sequence. The outlier frames which have spectral distortion greater than 2 dB are 3.63 %. A switched classifier is built to reduce the outlier frames down to about 0.27 % and make no frames with spectral distortion greater than 4 dB at an average bit-rate below 19 bits per frame.
Bibliographic reference. Jean, Fu-Rong / Wang, Hsiao-Chuan (1994): "A two-stage coding of speech LSP parameters based on KLT transform and 2d-prediction", In ICSLP-1994, 2091-2094.