Elitist identification of stops from formant transitions
Abstract
We present a set of strong cues for stop place of articulation, based upon formant transition slopes. This set complements a first one defined from stop bursts. Strong cues, which identify or eliminate a phonetic feature with certainty (no error is allowed), are not systematically fired. They can be exploited in language learning (cue enhancement) or in ASR to reduce the search space during the lexical access. The detection of the formant transitions and the cues is entirely automatic. To that purpose, we elaborate a robust formant tracking detector able to follow the transition trajectory. We obtained an average firing rate of 30 % on a clean corpus made up of sentences.