Faster time-aligned phonetic transcriptions through partial automation

A semi-automatic process for generating time-aligned transcriptions of speech data at the word and phone level is described. At each stage in the process, segment durations are estimated to generate approximate boundary markers, which are then corrected by hand. Corrections at one level are taken into account in the generation of boundaries for the next level, such that the error is reduced at each successive stage. A test implementation based on Praat was applied to a corpus of Brazilian Portuguese and a comparison against a fully manual process revealed a reduction of 54% in the time required to generate phonetic transcriptions and an average error of 21 ms in the time-alignment of phonetic boundaries.