Evaluation of cepstrally derived harmonics-to-noise ratios in voice signals

The harmonics-to-noise ratio (HNR) has been shown to provide a useful, non-invasive index for assessing and analyzing voice disorders. As such, a valid and reliable method for estimating levels of noise and hence HNR, in human voice is important for effective evaluation and management of voice pathology. An objective method to determine the accuracy of two cepstrum-based techniques described in the literature for calculating the harmonics to noise-ratio (HNR) in voiced speech is proposed. In order to draw an objective comparison between the two HNR estimation techniques, the same analysis parameters (analysis window size, voice source parameters (same open and closed quotient), same level of noise for each synthetic speech signal being analyzed) are used. A reference benchmark is defined for HNR estimation comparison as the ratio of the clean signal energy to the noise energy expressed in decibel. This reference HNR forms the basis upon which the two HNR techniques are evaluated for accuracy. Accuracy of a method is measured in terms of deviation of the estimated HNR from the reference HNR for a given noise level. A comprehensive explanation of the origins of the differences in accuracy of each approach is presented. Fig.1. HNR estimation using de Krom [1] cepstral baseline technique using 1024, 2048, 4096 window lengths (1024 shown).