Modelpublic

Rayantion26/JINGSI-Taiyu-STT-Preview

sourceHugging Faceotherupdated 19d agoView on Hugging Face

0likes31downloads

Model Card

JINGSI Taiyu STT Preview

A rank-8 LoRA/PEFT adapter for MERaLiON/MERaLiON-3-3B-ASR, adapted as a provisional Taiwan-Taiyu subtitle-target ASR preview.

Status — presentation preview only

—Supervision: project-owner-designated, authorized burned-in Taiyu subtitles extracted with local OCR.
—Training: 346 duration-filtered 16 kHz clips from one authorized source video.
—Validation: 629 clips from a separate source video; 50-clip subtitle-target mean loss: 3.0681.
—Qualitative sample: mixed results, including exact and partial matches.
—No native-speaker-reviewed Taiyu WER/CER exists.
—This is an editable machine-draft preview, not reliable unedited transcription.

The adapter contains no video, source audio, subtitle corpus, OCR crops, held-out references, or credentials.

Load

python

import torch
from peft import PeftModel
from transformers import AutoModelForSpeechSeq2Seq, BitsAndBytesConfig

base = AutoModelForSpeechSeq2Seq.from_pretrained(
    "MERaLiON/MERaLiON-3-3B-ASR",
    trust_remote_code=True,
    quantization_config=BitsAndBytesConfig(load_in_4bit=True),
    device_map={"": 0},
)
model = PeftModel.from_pretrained(base, "Rayantion26/JINGSI-Taiyu-STT-Preview")

Use the matching MERaLiON processor/prompt format and 16 kHz mono audio.

Limitations

OCR may confuse Traditional/Simplified glyph variants and captions are the project-designated target, not independently verified verbatim speech. Do not use this preview for high-stakes, legal, medical, safety, or unreviewed transcription.

Attribution and licence

This is a modified derivative adapter for MERaLiON, developed by the Agency for Science, Technology and Research (A*STAR), Singapore. See NOTICE and LICENSES/MERaLiON-3-Public-Licence.pdf. Gemma is provided under and subject to the Gemma Terms of Use at ai.google.dev/gemma/terms.