Analysis and Evaluation of Time-Domain Techniques for Pitch Extraction
Hiroya Fujisaki, Yoshihisa Tanabe · The Journal of the Acoustical Society of America · 1972
Recent techniques for pitch extraction are mostly based on detection of short-time periodicity of certain time functions, such as autocorrelation function or cepstrum, derived by some transformations of the original speech signal. No systematic studies, however, have been made for the analysis and quantitative evaluation of the performance of these techniques. This paper presents results of both theoretical and experimental investigations on this problem. Deterioration in the observed periodicity is caused partly by the nonstationarity of the signal itself, and partly by the finite length of the time window of observation as well as its relative phase to the excitation signal. These effects are theoretically estimated for a given set of parameters of the speech signal and the analysis procedure, and lead to quantitative evaluation and comparison of various techniques thus far proposed. Results of the theoretical analysis are then confirmed and supplemented by experimental analysis of performances of these techniques for natural speech.