Creating Small ASR Models For Low-resource Languages From Similar Languages: The Cases Of Betawi and Puno Quechua

This paper describes the creation of two small models for Betawi and Puno Quechua. To do so without additional training data in sufficient quantity, we rely on the data and models available in other closely related languages. For Betawi, we fine-tuned a Whisper-base model activated on several related languages: Malay, Javanese and Indonesian. The best results are achieved using the model fine-tuned on Malay. For Puno Quechua, we performed a three stage fine-tuning of the wav2vec2-base model by using data from several related Quechua variants and a small additional dataset in Puno Quechua. In order to improve the WER achieved by this model, particularly word segmentation, we also explored a probability-based in-context learning approach as a post-processing step to rerank the hypotheses from the wav2vec2 model.

HAL preprint (Mozilla Shared Task)