Cross-language Phoneme Mapping for Low-resource Languages: An Exploration of Benefits and Trade-offs

Voice-based systems are an essential approach for engaging directly with low-literate and underrepresented populations. Previous work has taken advantage of high-resource speech recognition technology for low-resource language speech recognition through cross-language phoneme mapping. Unfortunately, there is little guidance in how to deploy these systems across a range of languages. We present a systematic exploration of four source languages and five target languages to understand the trade-offs and performance of different source languages and training techniques. We find that one can improve recognition accuracy by selecting a source language that has similar linguistic properties to that of the target language. We also find that the number of alternative pronunciations per word and gender of participants also impact recognition accuracy. Our work will allow other researchers and practitioners to quickly develop high quality small-vocabulary speech-based applications for underresourced languages.