Pseudo-maximum-likelihood speech pitch extraction

The pitch of a voiced speech segment may be estimated by minimizing the squared error in approximating the segment by the repeated weighted average of successive assumed pitch periods, after a "maximum-likelihood" solution originally due to Slepian. A modification is proposed which allows use of an arbitrary window function and corrects two faults of the "likelihood" function noted by Noll: a linear baseline rise and ambiguous identification of the proper peak. Computationally, the entire formulation is in the time domain. Results are presented for speech both before and after bandpass filtering to simulate telephone transmission. Comparison of these two cases implies that the algorithm is effective for speech in which the fundamental frequency itself is missing.