Spectral widening of telephone speech using an extended classification approach

The quality degradation of transmitted speech in current telephone systems due to the band limitation can be reduced by application of an enhancement system which generates “pseudo” wide-band speech from its telephone-band version. In this contribution, an extended classification approach is presented as a component of such an enhancement system. The incoming telephone-band signal is divided into the spectral envelope and the residual signal by means of an LPC-error-filter. The proposed method focuses on the enhancement of the spectral envelope. At this, an appropriate set of spectral wide-band envelopes is selected from a pre-trained codebook in two steps. In the first step a subset of proper wide-band envelopes is selected. This is then further reduced in the second step by means of a cost-function which takes account of the relationship between the wide-band envelopes of the subset and the resulting wide-band envelope of the classification procedure one time step earlier. Listening tests have shown that the extended classification is a proper approach to enhance the quality of the synthesized signal components.