Formosan ASR — Saaroa

This is FormosanBank's current general-purpose automatic speech recognition checkpoint for Saaroa. It was selected by the lowest normalized micro word error rate among the completed language-level TACL systems evaluated on the same frozen all-corpora test partition.

Model

Repository: FormosanBank/formosan-asr-saaroa
Architecture family: XLS-R 53
Training path: train-only pan-Formosan continued pretraining, pan-Formosan supervised fine-tuning, and language adaptation
Base model: facebook/wav2vec2-large-xlsr-53
Parameters are stored in model.safetensors

Training lineage

Base checkpoint: facebook/wav2vec2-large-xlsr-53.
Continued self-supervised pretraining on the frozen pan-Formosan training partition only; development and test audio were excluded.
Supervised pan-Formosan ASR training, then adaptation to the target language's pooled training corpora.

The target-language supervised stage used ILRDF, ePark1, ePark2, ePark3. The experiment used dataset_v1, frozen split_v2, normalization formosan_safe_v1, and seed 13. Full machine-readable provenance is included in training_config.json.

Evaluation

Evaluation uses the frozen, leakage-controlled split_v2 test manifests and normalized text. Metrics are computed by summing edit counts across all available Saaroa test corpora, so larger corpora contribute in proportion to their reference tokens.

Normalized micro WER: 32.19%
Normalized micro CER: 4.22%
Test utterances: 1,014
Test corpora: 4
Automated readiness: review

The automated readiness label is review, with zero hard failures. The retained warning is recorded in readiness_summary.json and should be considered when reproducing or promoting the checkpoint.

Corpus	Utterances	WER	CER
ILRDF	414	42.19%	5.98%
ePark1	161	26.00%	2.53%
ePark2	94	33.80%	8.10%
ePark3	345	24.46%	1.92%

Detailed edit counts are provided in evaluation_results.json and metrics_by_corpus.tsv.

These figures are specific to the frozen TACL evaluation protocol. They should not be compared directly with the April 2026 stage-one releases, which used a different experiment snapshot.

Usage

from transformers import pipeline

repo_id = "FormosanBank/formosan-asr-saaroa"
transcribe = pipeline("automatic-speech-recognition", model=repo_id)
result = transcribe("path/to/16khz_audio.wav")
print(result["text"])

For long recordings, segment the audio before inference. The training recipe used utterances between 2 and 20 seconds.

Intended use

This checkpoint supports research, education, language documentation, and revitalization work involving Saaroa. Human review is recommended before using transcriptions in archives, publications, teaching materials, or other consequential settings.

Limitations

Performance varies substantially across corpora, speakers, recording conditions, dialects, speaking styles, and orthographic conventions. The aggregate score can hide weak performance on a particular corpus. The model may omit, substitute, or hallucinate words and should not be treated as an authoritative transcription source.

Data provenance

Training and evaluation artifacts come from the frozen Hunter Formosan TACL pipeline. Source corpus IDs can include Bible, ILRDF, NTU, ePark1, ePark2, ePark3, Xuan, YeddaPalemeqBlog, Youtube, and YutasWilang, depending on language availability. Audio is not redistributed in this repository. Underlying recordings retain their original rights and access conditions.

License and attribution

This model release and FormosanBank annotations/metadata are provided under CC BY 4.0. The upstream facebook/wav2vec2-large-xlsr-53 checkpoint is distributed under Apache 2.0. Users remain responsible for respecting the terms associated with underlying source recordings.

Please cite FormosanBank:

@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}
}