I like your tone: The effects of phonological similarity versus model training hours on cross-language forced alignment in tonal languages

Using intermediate languages to fill in the gaps of data for under-resourced languages is an important technique in automatic speech recognition (ASR). Cross-Language Forced Alignment (CLFA) uses acoustic modelstrained on one language to automatically generate audio-text alignment for another language. CLFA makes the forced alignment technique more accessible for low-resource languages. Discovering the factors that make CFLA more effective will improve this accessibility. Tonal languages are often under-resourced. However, little study has been done to investigate the factors that make CLFA more effective for tonal languages. The present study aims to answer this question by comparing CLFA alignment results from a higher-resource Cantonese acoustic model, a lower-resource Cantonese model acoustic model, and a lower-resourced Vietnamese acoustic model on aligning the same Thai audio data. The alignment results show that Cantonese model with 300 h achieve 54% accuracy, outperforming the Cantonese model with 10 h by 4%, but underperforming the Vietnamese model (∼100 training data) by 21%. Overall, our findings suggest that, although more training hours offer better CLFA performance for tonal languages, phonological similarity between languages (Thai and Vietnamese tonal similarity between languages) positively affected CLFA performance the most.