A time-frequency domain approach for pitch-peak picking of noisy speech

This paper presents an improved autocorrelation based algorithm for pitch estimation of noisy speech. The autocorrelation function of the pre-processed noisy speech signal is weighted by an inverse of its variant, circular average magnitude difference function (CAMDF). The inversed CAMDF-weighted autocorrelation function (ICWAF) alone cannot guarantee the de-emphasization of the `non-pitch-peaks'. However, as the dominant pitch-harmonic (DH) of the sinusoidal clean-speech model is expected to be an integer multiple of the pitch, this method concentrates in the effective estimation of the DH using a proposed cosine model of autocorrelation function. Hence, a variable period impulse train with its period being a function of the DH is optimally fitted with the rectified ICWAF to estimate the pitch. The simulation results using the Keele database show that the proposed method claims better accuracy even at a signal to noise ratio (SNR) as low as -10 dB.