Improvements in estimating the harmonics-to-noise ratio of the voice

An improved algorithm for computing harmonics-to-noise ratio (HNR) of the voice was developed and tested using synthesized vowels at various fundamental frequencies and known HNR values and showed mean absolute errors between computed versus expected HNR to be < 2 dB, a significant improvement over previously reported methods.

Abstract

An improved algorithm for computing harmonics-to-noise ratio (HNR) of the voice was developed and tested using synthesized vowels at various fundamental frequencies and known HNR values. The described algorithm implements a stringent routine for the identification of pitch pulse boundary markers based on a combination of peak-picking and zero-crossing procedures. Results showed mean absolute errors between computed versus expected HNR to be < 2 dB, a significant improvement over previously reported methods. Computational errors were minimized most effectively with vowel samples in the range of 17 to -10 dB HNR. Normative data for young adult males and females indicate that HNR < 13 dB may be considered abnormal in a majority of cases.