ISCA Archive - Statistical discrimination of French initial stops (original) (raw)

Statistical discrimination of French initial stops

S. Kitazawa, Jean-Pierre Tubach

Studies on the invariant features of Japanese stop consonants have been extended to French. Place and or manner of articulation of 7 consonants /?,p,t,k,b,d,g/ are discriminated in an environment of /a,o,ce,e,e:,u,y,i,a,e,6/. The feature vector is 23 LPC cepstrum coefficients at every 10ms of the initial 100ms (30ms before the burst and 70ms after the burst). The burst point was manually specified referring to waveform display. The stepwise discriminant analysis in the SAS system was used to obtain reduced feature set and discriminant score. The sample comprizes 3080 monosyllables from 40 male speakers. Speakers and vowel independent discrimination results better than for Japanese stops. The conclusion that the spectral pattern near the stop burst is a good feature for place discrimination can be generalized throughout French and Japanese. Results are compared with perception test.