Pitch estimation
Sarel van Vuuren · 1998
This report is about estimation of the pitch of a speaker. It is based on an earlier report by the author and is intended mainly to make the material presented there more accessible. It is intended as a review of a number of pitch estimation techniques with the objective being to highlight issues involved in obtaining a reasonable estimate of the pitch. These issues form the basis for a companion report where a technique using cepstral smoothing will be described. The first part of the report is a discussion of various pitch estimation algorithms and their relative strengths and weaknesses. The second part elaborates on a well-known technique that is based on the cepstrum. The cepstrum itself is typically derived from spectral information obtained using a short-time analysis window. Unfortunately, an apparent weakness of the cepstrum technique is that the length of this analysis window can affect the quality of the pitch estimate deleteriously depending on the pitch of the particular speaker. With the average pitch frequency of a speaker ranging from less than 80 Hz for a male, to more that 300 Hz for a female and as much as 500 Hz for a child, this can be a substantial problem. To alleviate this problem we propose a technique to adapt the length of the analysis window iteratively based on the estimated pitch of the speaker. We provide estimation results for speech from the TIMIT corpus.