Vowel recognition from harmonic spectra
James M. Hillenbrand, Robert A. Houde, Terrance M. Nearey, Michael J Clark · The Journal of the Acoustical Society of America · 1999
Vowel classification methods based on either formant representations or overall spectral shape nearly always begin with the calculation of a smoothed, pitch-independent spectrum. In this report, a method is described for the classification of vowels directly from high-resolution spectra that retain harmonics of the voice source. Recordings were made of ten sustained vowels spoken by 20 men and 20 women. Smoothed spectral-shape templates for each of the vowels from a random half of the talkers were constructed as the average of the normalized spectra of like vowels spoken by different talkers. The key steps in the normalization procedure consist of: (1) removal of spectral tilt by the subtraction of a running average of spectral amplitudes, (2) dynamic range compression of the tilt-adjusted spectrum, and (3) smoothing across frequency. High-resolution harmonic spectra from the remaining half of the talkers were then classified by computing the bin-for-bin difference between the harmonic spectrum and each of the smoothed spectral-shape templates.