Exponentiated enhancement for fundamental frequency extraction of noisy speech

The fundamental frequency extraction of speech is an important problem needed in many speech processing systems. In this paper, we propose a new method for extracting the fundamental frequency of noisy speech using correlation, difference, and exponent. The inverse Fourier transform of the exponentiated amplitude spectrum is weighted by the inverse of the average exponentiated magnitude difference function. By using an appropriate exponent constant determined from the noise amount, the signal component corresponding to the fundamental frequency is emphasized. We compare the proposed method with conventional ones. The simulation results obtained with the appropriate exponent at each signal-to-noise ratio show the effectiveness of the proposed method.