Evaluation of ASR front ends using synthetic vowel-like sounds
Hynek Heřmanský, Hector Raul Javkin · The Journal of the Acoustical Society of America · 1986
Analysis techniques and distance metrics interact in their effectiveness as the front end of the automatic speech recognizer (ASR). Synthetic vowel-like sounds were used to compare different ASR front ends, focusing on cepstrally smoothed standard LP analysis and the recently proposed perceptually based LP (PLP) analysis combined with the LP cepstral distance and the root power sum (RPS) distance metrics. (1) The findings of Kamm and Kahn [J. Acoust. Am. Suppl. 1 78, S82 (1985)] that standard LP overemphasizes the importance of higher formats were confirmed. The tonality scale inherent in PLP representation alleviates this problem. (2) In PLP, as opposed to LP, the contribution of a formant change depends much more on the particular vowel in which it occurs. (3) LP is less sensitive to changes in formant bandwidth than PLP. (4) RPS is less sensitive to the overall spectral slope changes than the cepstral metrics. (5) With PLP, RPS emphasizes the dominant formant more than the cepstral metric. (6) All examined techniques are sensitive to the harmonic structure of speech. The re-analysis of published formant frequency jnd experiments suggests this sensitivity might not be inconsistent with human perception.