Recursive forced alignment: A test on a minority language

Abstract

We compare recursive and linear approaches to force-aligned data from Matukar Panau, an endangered language of Papua New Guinea. Data were force aligned with the train/align procedure in the Montreal Forced Aligner. Using manual alignments produced by a trained phonetician as a benchmark, the recursive approach was found to outperform the linear approach. The recursive approach produced alignments that overlapped more with those made by human coders, and resulted in fewer fluctuations in both Overlap Rate and Error Rate. We conclude that a recursive approach enhances the quality of automated alignment of languages lacking a pre-existing acoustic model.