Using decision trees to construct optimal acoustic cues

This paper presents an approach to the optimization of acoustic cues used for stop identification in the context of an acoustic-phonetic decoding system which uses automatic acoustic event extractors (a formant tracking algorithm and a burst analyzer). The acoustic cues have been designed on the basis of acoustic studies on stops and spectrogram reading experiments. This ensures that these cues have a certain amount of discriminating power but we do not know either the optimal thresholds nor which combination of cues, are the most efficient. Therefore, we propose to use the decision tree theory to choose the most discriminating power and to improve their discrimination power. Considering the stop occurrences of a training corpus, the best cues are those which allow the decision tree leading to the best partition to be constructed. We have considered all the cues derived from the ones provided by the phonetician on formant transitions and burst characteristics. The improvement of the cues has been achieved on a corpus of 941 stops.