Multi-mode quantization of adjacent speech parameters using a low-complexity prediction scheme

This work addresses joint quantization of adjacent speech parameter values or vectors. The basic joint quantization scheme is improved by using a low-complexity predictor and by allowing the quantizer to operate in several modes. In addition, this paper introduces an efficient algorithm for training quantizers having the proposed structure. The algorithm is used for training a practical quantizer that is evaluated in the context of the quantization of the linear prediction coefficients. The simulation results indicate that the proposed quantizer clearly outperforms conventional quantizers both in an error-free environment and in erroneous conditions at all bit error rates included in the evaluation.