Computer Recognition of Phonemes in the Presence of Cockpit Induced Stress and Noise

摘要

Speech recognition algorithms were analyzed using normal and G-stressed speech as an input. Speech samples were recorded in centrifuge tests at the Air Force Medical Research Lab, Wright-Patterson AFB, Ohio. All speech was recorded using the MBU-12/P face mask. The algorithms studied are phoneme-based feature extractors which feed a recognition algorithm based on fuzzy set theory. Three feature extraction algorithm options were analyzed. One option used a phoneme length of 40 ms and the other options used a length of 8 ms. The recognition results for all three options using normal speech are above 90%, but the 40ms phoneme length give higher raw scores. For G-stressed speech the 40 ms phoneme length scored greater than 90% while the 8ms phoneme length options scored less than 60%. (Author)