Towards a probabilistic system for estimation of acoustic landmarks for speech recognition

Abstract

Algorithms for consonant landmark detection, such as Liu [S. Liu, J. Acoust. Soc. Am. 100, 3417–3430 (1996)], extract cues to specific types of abruptnesses in the acoustics. The abruptnesses indicate occurences of closure and release for obstruent and sonorant consonants, and burst release for stop consonants. In Liu’s algorithm, fixed thresholds are used to filter out abruptnesses that are unlikely to be true landmarks. The resulting set of landmarks does not retain any information regarding these filtered-out instances. However, such information may be useful later in the lexical access process, especially given the range of contextual variation in the speech signal. In this work, the landmark detection process is reformulated as a probabilistic system. First, thresholds are lowered to include more candidates, and then a probability value is calculated for each candidate. An N-best search is used to pick the most likely sequences of obstruent landmarks based on the calculated probabilities. Experiments with 80 sentences from the TIMIT database detect corresponding landmarks within 40-ms windows of 96% of hand-labeled obstruent landmarks, 98% of burst-release landmarks, and 76% of sonorant landmarks. Applying 5-best search results in 9% deletion and 4% insertion rate. [Work supported by NIH DC02978.]