Matthew C. Kelley, assistant professor of linguistics, recently published an article in Phonetica 81(5): 451-508, with co-authors Scott James Perry (University of Alberta) and Benjamin V. Tucker (Northern Arizona University). The article is titled "The Mason-Alberta Phonetic Segmenter: a forced alignment system based on deep neural networks and interpolation."
In the article, Kelley and colleagues described a new software system they developed called MAPS (the Mason-Alberta Phonetic Segmenter) that automatically aligns a transcription of what someone said and when it occurs in a recording. Their tool MAPS outperformed the current state-of-the-art system with a 28.13% relative increase in an important metric for speech science: how many alignments were within 10 milliseconds of where a human annotator would place them. They also developed a technique to increase the precision of the alignments.
MAPS was built using machine learning techniques with deep neural networks, a form of artificial intelligence. The software is currently available for processing American English data. It can be downloaded from the Mason PhonLab GitHub, where future features and improvements will be released.
October 09, 2024