A robust pitch estimation approach for colored noise-corrupted speech

We present an integrated pitch estimation approach for severely colored noise-corrupted speech. An effective colored noise-whitening process is first applied to the noisy speech. Then, a variable-length average magnitude difference function (VLAMDF) of the pre-filtered noisy speech (PFNS) is proposed, which almost conquers the trend of falling valleys in the conventional AMDF. The amplitude characteristic of the VLAMDF is reshaped by means of a simple linear transformation to reduce the possibility of double-pitch-errors. As the VLAMDF exhibits a valley while the autocorrelation function (ACF) of PFNS provides a peak, the ACF is weighted by the reciprocal of the VLAMDF to emphasize the pitch-candidate as well as to suppress the non-pitch peaks. Moreover, a noise-robust pitch detection in the time-domain is guaranteed by collaboration of this enhanced autocorrelation function with the reshaped version of the VLAMDF. The proposed approach is simulated using the Keele reference database and provides a superior accuracy relative to some of the existing methods implemented in the presence of colored noise, even at a very low signal-to noise ratio (SNR) of -15 dB.