Which cross-language acoustic model to choose: Assessing tonality and phonemic inventories effects

Cross-language forced alignment techniques, i.e., using a high-resource language acoustic model to automatically align audio files with transcripts for another language, present a promising and feasible approach for expediting the acoustic analysis of low-resource languages (Chodroff, Ahn, and Dolatian 2024). However, the factors influencing the effectiveness of cross-language forced alignment remain underexplored. Mandarin and Japanese share more similarities in phonemic inventories and syllable structures, while Japanese and English are both non-tonal languages. Comparing the alignment results generated by Mandarin and English acoustic models on Japanese audio data can help us explore whether tonality and phonemic inventory similarity have an equal influence on cross-language alignment performance. The present study used 30,975 Japanese audio files from Common Voice datasets as input and Montreal Forced Aligner Mandarin, English, and Japanese acoustic models to generate forced alignment output. In the 1.5 million data points across different comparison pairs, we found that when performing a cross-language alignment on a non-tonal language, a non-tonal language acoustic model is the optimal choice, but a tonal language acoustic model could also work decently if assigning vowels to a falling tone. These findings suggest tone matters in cross-language alignment and offer methodological insights for implementing cross-language alignment in low-resource languages.