Evaluating cross-linguistic forced alignment of conversational data in north Australian Kriol, an under-resourced language

Abstract

Speech technology is transforming language documentation; acoustic models trained on “small” languages are now technically feasible. At the same time, forced alignment built for major world languages has matured and now offers ease of use through web interfaces requiring low technical expertise. This paper provides an updated and detailed evaluation of cross-linguistic forced alignment, the approach of using forced aligners untrained on the target language. We compare two options within MAUS (Munich Automatic Segmentation System): language-independent mode vs major world language system (here, Italian) on the one dataset, a comparison that has not previously been reported. The dataset comes from a corpus of adult conversational speech in Kriol, an English-based creole of northern Australia. The results of using MAUS Italian were better than those of using the language-independent mode and those in previous studies: the agreement rate at 20 ms was 72.1% at vowel onset and 57.2% at vowel offset. With completely misaligned tokens excluded, the overall agreement rate rose to 69.2% at 20 ms and over 90% at 50 ms. Most errors in the output SAMPA (Speech Assessment Methods Phonetic Alphabet) labels were resolvable with simple text replacements. These results offer updated benchmark data for an untrained, late-model forced alignment system.