OpenVoiceOS/whisper-small-nan-tw-pinyin-onnx

whisper-small-nan-tw-pinyin — ONNX

ONNX export of linshoufan/linshoufanfork-whisper-small-nan-tw-pinyin (Whisper small fine-tuned for Taiwanese Hokkien/Min Nan, output in Tâi-lô romanization, by linshoufan) for onnx-asr (standard whisper model type — works with stock onnx-asr, no patches needed). fp32 and int8 variants included.

License: apache-2.0, inherited from the source model.

First specialized ONNX ASR model for Min Nan / Taiwanese Hokkien in this collection.

Usage

Output is romanized Tâi-lô, not Han characters. Whisper has no dedicated nan language token; use language="zh" — the Min Nan / romanization behavior comes from the fine-tune itself.

import onnx_asr
model = onnx_asr.load_model("whisper", "path/to/this/repo")  # or quantization="int8"
print(model.recognize("audio_16khz.wav", language="zh"))

Verified on a clip from sarahwei/Taiwanese-Minnan-Sutiau (a Hokkien word/phrase dictionary corpus; FLEURS does not cover Min Nan):

• Reference: 二步七仔 (jī-pōo-tshit-á)
• fp32: "jī-pōo-tshit-á" — exact match (RTF 0.98)
• int8: "jī-pōo-tshit-á" — exact match (RTF 0.64)

Both precisions reproduce the Tâi-lô romanization exactly. RTF measured on an AMD Ryzen 5 7600 (CPU, 4 OMP threads, shared/loaded box — not a clean benchmark number; note this clip is short, so RTF is noisier than on longer clips used for other languages in this collection).

Int8 decoder was produced by quantizing the pre-merge decoders separately and re-merging (merge_decoders(..., strict=False)); direct quantization of the merged decoder graph does not shrink it (its If subgraphs are skipped by onnxruntime's dynamic quantizer).