A robust pitch detection algorithm for speech signals in a practical noisy environment

Abstract

In this paper, a robust pitch detection algorithm is proposed for speech signals severely corrupted by a non-stationary noise. Using low-frequency band of noisy speech, an effective noise reduction approach is first formulated for power spectral subtraction to track the time-variation of the non-stationary noise prior to pitch detection. Then, a new normalized circular difference function of the enhanced speech, which almost conquers the constraint of overlapping between the first formant and the pitch, is proposed. Simulation results using the Keele reference database demonstrate a better efficacy of the proposed algorithm relative to some of the existing methods in a practical multi-talker babble noise environment.