Gaussian Mixture Model-based Quantization of Line Spectral Frequencies for Adaptive Multirate Speech Codec

In this paper, we investigate the use of a Gaussian Mixture Model (GMM)-based quantizer for quantization of the Line Spectral Frequencies (LSFs) in the Adaptive Multi-Rate (AMR) speech codec. We estimate the parametric GMM model of the probability density function (pdf) for the prediction error (residual) of mean-removed LSF parameters that are used in the AMR codec for speech spectral envelope representation. The studied GMM-based quantizer is based on transform coding using Karhunen-Loève transform (KLT) and transform domain scalar quantizers (SQ) individually designed for each Gaussianmixture. We have investigated the applicability of such a quantization scheme in the existing AMR codec by solely replacing the AMR LSF quantization algorithm segment. The main novelty in this paper lies in applying and adapting the entropy constrained (EC) coding for fixed-rate scalar quantization of transformed residuals thereby allowing for better adaptation to the local statistics of the source. We study and evaluate the compression efficiency, computational complexity and memory requirements of the proposed algorithm. Experimental results show that the GMM-based EC quantizer provides better rate/distortion performance than the quantization schemes used in the referent AMR codec by saving up to 7.32 bits/frame at much lower rate-independent computational complexity and memory requirements.