Quantization noise modeling in low-delay speech coding

We present a low-delay vector predictive transform coder with quantization noise modeling. The coder uses overlapped transform coding and a Kalman filter with a backward estimated LPC model of the speech signal combined with an additive noise model of the quantization noise. The noise model is driven by the vector quantizer. Each quantizer cell has its own error correlation matrix, which is kept in a table and addressed using the quantizer index. Simulation results indicate that overlapped transform coding combined with noise modeling can improve the quality of the decoded speech signal