Improved a priori SNR estimation with application in Log-MMSE speech estimation

A speech enhancement method utilizing the harmonic structure of speech is presented. The method is an extension of the well known minimum mean square error log-spectral amplitude estimator (Log MMSE) method for speech enhancement. The improvement lies specifically on a priori SNR estimation by utilizing harmonic structure of speech. The method is based on a conditional averaging operation over adjacent frequency bands for each processed data block. The actual frequency bands used in the conditional averaging is determined by a pitch detector. Thus voiced segments are averaged over frequency according to the pitch and the corresponding harmonic structure of voiced speech. Non-voiced segments are averaged over frequency according to a random number depending on the pitch value. The result is overall better SNR and SNRSeg values in white noise over the standard Log MMSE reference method. In babble noise, the estimator rendered similar SNR and SNRSeg values as the Log-MMSE reference method. Subjectively the residue background noise sounded more natural when using the suggested method.