A fast frequency domain pitch algorithm
Richard A. Gillmann · The Journal of the Acoustical Society of America · 1975
An algorithm for the extraction of fundamental frequency (pitch) from digitized speech is described. The computer algorithm operates in the frequency domain using autocorrelation spectra and is optimized for computational efficiency. Three basic operations are performed: (1) the digitized speech is low-pass filtered and down-sampled to a rate of 2000 samples per second, thereby eliminating frequencies outside the range of possible fundamentals and reducing the quantity of data to be processed. (2) Autocorrelation spectra are computed every 10 msec and pitch values are extracted. (3) The values of fundamental frequency obtained above are edited to eliminate anomalies. Reliability in terms of making the correct voiced/unvoiced decision exceeds 99%, and the program runs in about 14 times real time on a 16-bit minicomputer.