Speech Bandwidth Extension Based on Codebook Mapping and GMM

Speech bandwidth extension (BWE) based on the conventional Gaussian mixture model (GMM) often suffers from the overly smoothed problem, and the main reason is the low accuracy of the estimated covariance which results in the loss of specific high frequency feature. Thus, a speech bandwidth extension base on codebook mapping (CM) and GMM was proposed in this paper. Firstly, the feature of low frequency (LF) and high frequency (HF) were extracted, and the GMM model was trained. Then, an offset vector codebook was designed based on the trained GMM parameters. In the reconstruction phase, LF offset vectors were transformed to HF offset vectors according to the trained offset vector codebook. The final HF feature parameter was obtained by adding the HF offset vectors to the estimated part by GMM. It is shown by subjective evaluations and objective evaluations that the CM-GMM significantly overcomes the overly smoothed problem and obviously improves the quality of the synthesized speech signals compared with the conventional GMM-based BWE method.