Automating phonetic measurement: The case of voice onset time

We present an architecture for locating phonetic events accurately in time, and for measuring time differences between nearby events, using Voice Onset Time (VOT) as a case study. Although VOT remains a central concern in the field, phoneticians' VOT measurements generally continue to rely on human judgment. This requires significant labor, makes even large laboratory experiments onerous, and prevents the field from taking full advantage of the millions of hours of digital speech now becoming available. Our algorithm accurately automates VOT measurement, by combining HMM forced alignment for determining approximate stop boundaries with paired burst and voicing onset detectors. Each detector is a frame-level max margin classifier operating on the scale-space projection of a small number of relevant acoustic features. On a large set of clean lab speech, this system has a mean absolute error (relative to human annotation) of only 2.8 ms, with 98% of errors <10 ms. On a subcorpus independently annotated by two of the authors, the system agreed with the two human annotators as well as they agreed with one another (1.49 ms vs 1.50 ms). Promising results on other datasets will be reported. The system will be released as open-source software.