Audio Signal Processing and Coding

7.10 MDCT WITH VECTOR QUANTIZATION

Iwakami et al. developed transform-domain weighted interleave vector quantization (TWIN-VQ), an MDCT-based coder that also involves transform coefficient VQ [Iwak95]. This algorithm exploits LPC analysis, spectral interframe redundancy, and interleaved VQ.

At the encoder (Figure 7.4.), each frame of MDCT coefficients is first divided by the corresponding elements of the LPC spectral envelope, resulting in a spectrally flattened quotient (residual) sequence. This procedure flattens the MDCT envelope but does not affect the fine structure. The next step, therefore, divides the first step residual by a predicted fine structure envelope. This predicted fine structure envelope is computed as a weighted sum of three previous quantized fine structure envelopes, i.e., using backward prediction. Interleaved VQ is applied to the normalized second step residual. The interleaved VQ vectors are structured in the following way. Each N-sample normalized second step residual vector is split into K subvectors, each containing N/K coefficients. Second step residuals from the N-sample vector are interleaved in the K subvectors such that the i-th subvector contains elements i + nK, where n = 0, 1, … , (N/K) − 1. Perceptual weighting is also incorporated by weighting each subvector by a nonlinearly transformed version of its corresponding LPC envelope component prior to the codebook search. VQ indices are transmitted to the receiver. Side information consists of VQ ...