In this paper several approaches of time and frequency domain-based algorithms to estimate harmonics-to-noise ratios (HNR) in voice signals are compared. The approaches covered incorporate a recent time-domain correction to a classic method, as well as a frequency-domain adjustment introduced here. The experimental comparisons include the number of pitch periods needed to obtain the best HNR estimates, as well as the sensitivity of the methods to different perturbations of the periodicity pattern, like shimmer, jitter, noise and combinations of them. Time domain methods show better performance than frequency-based approaches, and moreover, the correction to the ensemble-average time domain technique reduces the required number of pulses by an order of magnitude.
展开▼