Evaluating full automation of formant extraction in the German Plapper Corpus

Authors: Robert Fromont, Jennifer Hay, Daniel Duran, Allie Osborne, Melanie Weirich, Miriam Oschkinat, Stefanie Jannedy

Abstract:

We have collected over thirteen thousand recordings of German sentence readings elicited via the Plapper app as part of a citizen science project. However, manual correction of transcripts and phone alignments is a bottleneck for sociolinguistic study of these recordings, with less than 10% of recordings processed so far, as data collection continues. We evaluate the possibility of fully automating the transcription and alignment of the Plapper corpus. Two automated transcription processes were tried and their target-word error rates compared; BAS ASR Web Service, and using the text prompt read by the participants as the transcript. Furthermore two forced alignment systems were evaluated: WebMAUS and the Montreal Forced Aligner. Their outputs were compared with manually corrected alignments, using both overlap rate to compare them temporally, and correlation of resulting formant measurements. We found that using prompt text as a transcript is more accurate than ASR, and that both forced aligners produced sufficiently trustworthy alignments, with resulting formant measurements that correlated highly with those generated from manually corrected alignments.