Pitch estimation using mean shift algorithm on multitaper spectrum of noisy speech

In this paper, we apply the mean shift algorithm to estimate pitches of noisy speech from its multitaper spectrum. The noisy speech is first transformed into the multitaper spectrum, which can reduce the stationary noise. The pitch is extracted from the multitaper spectrum using the mean shift algorithm. After all estimates are collected, dynamic programming is used to obtain a smoothed pitch contour. We compare the performance of our method with two well-known algorithms on the Keele pitch database and demonstrate that it performs much well even at SNR as low as -15dB.