Fundamental Frequency Estimation of Low-quality Electroglottographic Signals

Abstract

Fundamental frequency (fo) is often estimated based on electroglottographic (EGG) signals. Because of the nature of the method, the quality of EGG signals may be impaired by certain features like amplitude or baseline drifts, mains hum, or noise. The potential adverse effects of these factors on fo estimation have to date not been investigated. Here, the performance of 13 algorithms for estimating fo was tested, based on 147 synthesized EGG signals with varying degrees of signal quality deterioration. Algorithm performance was assessed through the standard deviation σfo of the difference between known and estimated fo data, expressed in octaves. With very few exceptions, simulated mains hum, and amplitude and baseline drifts did not influence fo results, even though some algorithms consistently outperformed others. When increasing either cycle-to-cycle fo variation or the degree of subharmonics, the SIGMA algorithm had the best performance (max. σfo = 0.04). That algorithm was, however, more easily disturbed by typical EGG equipment noise, whereas the NDF and Praat's auto-correlation algorithms performed best in this category (σfo = 0.01). These results suggest that the algorithm for fo estimation of EGG signals needs to be selected specifically for each particular data set. Overall, estimated fo data should be interpreted with care.