Quantization noise spectral shaping in instantaneous coding of spectrally unbalanced speech signals

Abstract

In the context of centralized spectral equalization of speech in a telephone network, the signal is spectrally strongly unbalanced at the output of the equalizer, before being quantized, which results in low SNR at the receiver. We propose and evaluate experimentally two methods to reshape the quantization noise, in order to make it less perceptible in reception. The first one consists in finding the most probable quantization sequence, given the desired noise spectrum. In the second one, the filtered quantization error is added to the signal to be quantized.