Third International Conference on Spoken Language Processing (ICSLP 94)
A problem of long-standing interest in speech processing concerns the most appropriate representation for recognition purposes. The objective of this paper is to compare 6 acoustic and 4 articulatory representations in a task of determination of place of articulation of inter-vocalic plosives. Place of articulation recognition results were obtained based on linear discriminant analysis with the "jackknife" method in which the tokens from each individual are successively removed from the training set, and used as a test set. Systematic comparisons were performed under 3 different sets of conditions depending on whether or not the information about the end of the transition, the transition and the stable part of the vowel are integrated. The LPC cepstrum and two articulatory representations (DRM and Maeda's model) achieved the best recognition rate (86%). However, the 2 articulatory representations appeared to be more stable in terms of inter-speaker variability. The performances of the 7 others representations were found to be significantly lower (74% for formants, and 60% for LPC area).
Bibliographic reference. Soquet, Alain / Saerens, Marco (1994): "A comparison of different acoustic and articulatory representations for the determination of place of articulation of plosives", In ICSLP-1994, 1643-1646.