Three-channel time-domain pitch detector using different non-linear digital filters
Wolfgang J. Hess · The Journal of the Acoustical Society of America · 1979
Nonlinear distortion is an adequate means to enhance the first harmonic of speech signals for time-domain pitch determination. To evaluate the usefulness of nonlinear functions (NLFs) in pitch extractors, 16 NLFs were investigated: odd functions including linear processing, even functions (e.g., full-wave rectification), “mixed” functions (e.g., half-wave rectification), and rectified single sideband (SSB) modulation. These functions were applied to three categories of signals; undistorted male speech, band-limited male speech, and high-pitched female speech. No single NLF was found to perform well for all signal categories. Simultaneous application of three NLFs (odd, even, and SSB) with equal amplitude characteristics, however, gives good results. Using these NLFs, a three-channel time-domain pitch extractor is implemented. After application of the NLFs, the partial signals in the individual channels are linearly filtered by a crude approximation of the inverse filter, and preliminary period boundaries (markers) are derived by threshold analysis. At least in one channel a strong first harmonic and thus a correct marker sequence can be expected. The algorithm finally selects this channel by checking the regularity of the obtained markers.