On the estimation of pitch of noisy speech based on time and frequency domain representations

In this paper, we propose a new algorithm for pitch estimation from speech signals heavily degraded by additive noise based on both time and frequency domain representations. A least-squares minimization technique is first developed for the accurate estimation of a pitch-harmonic (PH) wherein a harmonic sinusoidal model of clean speech is exploited as a time domain representation. Then, relying on a power spectrum in the Fast Fourier Transform domain which is a frequency domain representation, a two-step criterion is formulated in order to acquire a true harmonic number corresponding to the extracted PH for robust pitch detection. Extensive simulations have been carried out to demonstrate the effectiveness of the proposed methodology as compared to some of the existing techniques in literature. It has been shown that our new approach consistently outperforms the other methods especially at low levels of signal-to-noise ratio (SNR).