A Study Of Bandwidth Extension Algorithms Based On GMM

The frequency of current telephone speech ranges from300Hz to3400Hz. When thebandwidth extended to300Hz-8000Hz, the quality of wideband speech sounds more nature andintelligible. With the development of signal processing technology, the demand of high qualityspeech is more and more widely. However, it’s impossible to fully updated the current network ina short period because of economics cost, so artificial bandwidth extension was proposed anddeveloped quickly. In the transition time, bandwidth extension has important significance. Thepaper proposed two improved algorithms in bandwidth extension, compared with traditionalbandwidth extension based on Gaussian mixture model.In order to avoid over-smoothing of the evaluated high frequency features based onGaussian mixture model in bandwidth extension process, the paper presents a method ofspeech bandwidth extension based on combination of self organizing featuremap and Gaussian mixture model (SOFM-GMM). In the training process, training data areunsupervised clustered by self organizing feature mapping network, then data which relatedhighly are divided into different clusters. After the training process, a GMM model isestablished for each kind of cluster. So every model can accurately represent relations betweenhighly similarity data.The reconstructing high frequency lost because of low accuracy of the establishedcovariance based on GMM, especially the covariance matrix based on GMM is a full matrixrather than a diagonal matrix. To resolve the problem, the paper presents a method of speechbandwidth extension based on combination of Codebook Mapping and and GMM. Codebooktraining is based on the parameters of GMM and shift vector. High feature parameters evaluatedby Codebook mapping is part of the final high feature parameters. Sum of the mapped shiftvector applied with a adjustment coefficient and estimated parameters based on GMM are thefinal feature parameters.At last, simulation experiments carried out. Both subjective evaluation tests and objectiveevaluation tests show that the wideband speech reconstructed by two proposed algorithmsounds better and more nature than that based on traditional GMM.