Robust pitch estimation using BPFP and NN in noisy environments

摘要

In this work, we study a pitch extraction method with U/V classification using bank of bandpass filter pairs (BPFP), or combining bank of BPFP with neural network (BPFP-NN), in performance with respect to addition of white/pink noise, as compared with modified autocorrelation (MAUT) analysis, when a threshold by which plus, zero or minus slope of frequency direction in power spectrum is ternarized is introduced into BPFP as measures against the influence of noise. The results of experiments show that BPFP and BPFP-NN methods introduced a threshold into BPFP are by no means inferior to MAUT method in noisy environments.