5th International Conference on Spoken Language Processing
A novel method for robust pitch extraction, based on the correlogram output of the Lyon's cochlear model is described. The value of the autocorrelation lag for which the signals of the cochlear channels have the same periodicity can be computed thus tracking how the pitch of the input signal varies in the time domain. In the case of a stationary noise, a sort of 'spectral-subtraction' technique, built in the correlogram domain named 'correlogram subtraction', is applied to enhance the signal before computing its fundamental frequency. Finally, a correction algorithm based on an 'island driven' strategy, working on particular zones of the signal with stable pitch values, is used to refine the pitch estimate. This method of pitch extraction is extremely reliable, even in the case of a signal to noise ratio of 0dB. The same subtraction technique, with some new specific filter-bank energy-based modifications, is considered to re-synthesize, by an inversion strategy, a clean version of an input noisy signal. The quality of the re-synthesized signal is quite promising, leading us to try, in the future, to use this technique as a new signal enhancement scheme.
Sound Examples: Italian word /lavan'daja/ ('washerwoman)
#1 - pronounced by a male speaker in clean condition
#2 - pronounced by a male speaker in a noisy condition (0dB SNR)
#3 - reconstructed, by the correlogram subtraction technique described in the paper, from the corresponding noisy signal (0dB SNR)
Bibliographic reference. Cosi, Piero / Pasquin, Stefano / Zovato, Enrico (1998): "Auditory modeling techniques for robust pitch extraction and noise reduction", In ICSLP-1998, paper 1053.