An exemplar-based approach to automatic burst detection in spontaneous speech

This paper introduces a novel algorithm for detecting burst in voiceless stops in spontaneous speech. This algorithm uses an exemplar-based approach for detecting aspiration noise, and avoids the normalization problem since the exemplars are inherently speaker-specific and environment-specific. The algorithm is trained and tested on 19 speakers’ data. The overall error is estimated to be under 5 ms. We also show the wide range of variation in the phonetic makeup of stops in spontaneous speech and how the algorithm is improved to deal with the difficult cases.