A robust pitch detection in noisy speech with band-pass filtering on modulation spectra

In this report, we are presenting new robust pitch detection for noisy speech. The conventional method, i.e., AUTOC, is vulnerable to the serious noise environment, especially the periodical noise. In the case of additive car noise, the detection accuracy is considerably deteriorated. A new detection method is proposed by adding a process, which implements band-pass filtering on the modulation spectra of the speech sections to AUTOC. The 2-nd power amplitude spectrum of speech in the autocorrelation computation of AUTOC is replaced by the 3-rd power amplitude spectrum. In addition, a band-limitation operation in frequency domain is carried out. It is adapted to the pitch features of human speech. An evaluation using 10 Chinese words is undertaken to compare the proposed detection method with AUTOC and a recent method based on exponentiated band-limited amplitude spectrum. The experiment is at the noise level ranged from 0 dB SNR to 10 dB SNR with white noise, colored noise and car interior noise. It is shown that the error of gross pitch error (GPE) in the proposed detection method is significantly decreased in severely noisy speech.