Stop identification using hidden Markov models

A series of experiments has been undertaken to assess the power of discrete spectral slices for automatically discriminating between the voiceless plosives /p,t,k/ in CV syllables. These experiments involve a 936-token data base consisting of 52 instances of each of the 18 syllables /p,t,k/ × /i,e,ae,a,ow,u/ spoken by 13 male and 13 female talkers. In one experiment, identification was attempted using a single 12-pole LPC onset spectrum. The onset spectra from each talker were compared, using a log-likelihood distance measure, with the 900 onset spectra of the remaining 25 talkers. An overall classification accuracy of 92% was achieved using a k-nearest-neighbor decision strategy. A second experiment involves classifiers which use a series of LPC spectra computed during the first 50 ms of the stop release. Each CV syllable is modeled as a hidden Markov process which generates a spectrum every 5 ms. Classification is performed using either a Viterbi or forward-backward decoding strategy.