Description:
<jats:p>The recognition accuracy for intervocalic voiced plosives using a simple consonant detector combined with a context-dependent Bayesian classifier was studied. The consonantal segments were located by running a dip detector over the smoothed log-energy contour. The vocalic nuclei were located at the segments of least spectral change. LPC spectra along the vocalic segments were the features for their classification which was performed by a compound Bayes decision procedure. The feature set for the Bayesian context-dependent classification of voiced stop sounds was a sequence of three selected LPC spectra located along the amplitude dip. One spectrum was obtained at the point of minimal energy and the other two were obtained at the points of maximum slope of the smoothed log-energy contour at the beginning and end of the consonantal segment. A recognition performance of 92% was obtained in a speaker-dependent manner using a database consisting of nearly 1300 consonants embedded in VCVCVCV nonsense sequences where the V's were the vowels /a,i,u/ and the C's were the voiced stops /b,d,g/ uttered by two male Argentine Spanish speakers.</jats:p>