FormosanBank/xls-r-53-stage1-kanakanavu-asr

Model description

This repository contains a FormosanBank automatic speech recognition model for Kanakanavu, fine-tuned from facebook/wav2vec2-large-xlsr-53.

It is part of the broader FormosanBank effort to support research, education, and revitalization for Taiwan’s Indigenous Formosan languages.

Training and release notes

Source experiment directory: /scratch/scheppat/formosan_asr/experiments/xlsr53_stage1_lang_finetune_basemodel/Kanakanavu/20260210-124835_2000411
Release contents were curated from the run's final/ directory plus processor artifacts.
This release is intended for inference and reproducible loading from the Hugging Face Hub.

Evaluation

Validation metrics from eval_results.json:

WER: 0.28618331053351576
CER: 0.16783158247185367
Epoch: 48.309178743961354

Usage

from transformers import AutoModelForCTC, AutoProcessor, pipeline

repo_id = "FormosanBank/xls-r-53-stage1-kanakanavu-asr"
processor = AutoProcessor.from_pretrained(repo_id)
model = AutoModelForCTC.from_pretrained(repo_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
)

result = pipe("path/to/audio.wav")
print(result["text"])

Limitations

This model is intended for Formosan language ASR research and educational use. Performance can vary by corpus, speaker, recording quality, orthography conventions, and domain mismatch.

FormosanBank

Project docs: https://ai4commsci.gitbook.io/formosanbank

Citation

If you use this model, please cite FormosanBank itself:

@misc{mohamed2024formosanbank,
  author       = {Mohamed, W. and Le Ferrand, É. and Sung, L.-M. and Prud'hommeaux, E. and Hartshorne, J. K.},
  title        = {FormosanBank},
  year         = {2024},
  note         = {Electronic Resource},
  url          = {https://ai4commsci.gitbook.io/formosanbank}
}

License and attribution

FormosanBank annotations and metadata are licensed under CC-BY-4.0. You must cite the source in any redistributed or derived products. For code packages, you may refer to the GitHub repository. For academic publications, cite the FormosanBank electronic resource above.