Downloads · 30 days
0
CHOCHICHO/twirly-whisper-large-v3-turbo-int8
twirly-whisper-large-v3-turbo-int8 is a automatic speech recognition model from CHOCHICHO. Use it when you need speech turned into text. It is set up for sherpa-onnx. The card lists the license as mit.
openai/whisper-large-v3-turbo exported to ONNX and dynamically quantized to int8, packaged for sherpa-onnx. No fine-tune — these are the unmodified OpenAI weights.
Downloads · 30 days
0
Access
Public
Updated Aug 6, 2026
Repo size
2.1 GB
Likes
0
Public
Click a slice to open those files.
.onnx1 GB · 100%
From the Hugging Face model README
openai/whisper-large-v3-turbo exported to ONNX and dynamically quantized to int8,
packaged for sherpa-onnx. No fine-tune —
these are the unmodified OpenAI weights.
Used by Twirly for Swiss German dictation on Android: the model hears Swiss German and writes Standard German.
| File | Bytes | SHA-256 |
|---|---|---|
encoder.int8.onnx | 676 517 811 | ca6d11c9c48163fc5643743b86a70751536ef3cbb9f3d4b87bff6b1fec9be302 |
decoder.int8.onnx | 361 352 573 | 3724fb420901b37e1015d99e10efa78e6df6bd3831c7889fcb9db2d81fe6ca31 |
tokens.txt | 816 730 | b34b360dbb493e781e479794586d661700670d65564001f23024971d1f2fa126 |
export-onnx.py from sherpa-onnx, with one patch: dynamo=False in both
torch.onnx.export calls. With the Torch 2.x default (dynamo=True) the graph loads
in sherpa-onnx without complaint, burns compute — and returns empty text.
python export-onnx.py --model turbo
Quantization: dynamic, op_types_to_quantize=["MatMul"], QuantType.QInt8, and —
unlike the script's default — per_channel=True, reduce_range=True. That is worth
1.2 WER points on the measurement below, for 2 MB more file.
392 segments of the Swiss Parliaments Corpus (CC BY 4.0), 11 270 reference words, word error rate over difflib opcodes:
| Weights | Decoding | WER |
|---|---|---|
| fp16, MLX | Whisper's full decoding | 19.9 % |
| fp16, MLX | greedy | 21.0 % (first 100 segments; full decoding 21.1 % on the same subset) |
| int8 per-tensor, sherpa-onnx | greedy | 24.6 % |
| int8 per-channel, sherpa-onnx (these files) | greedy | 23.5 % |
So the remaining ~3.6-point gap is the quantization, not sherpa-onnx's greedy search — the two decoding strategies land within 0.1 points of each other on identical weights.
⚠ Per-channel looked worthless on a 50-segment subset (30.5 % vs 30.4 %) and only separated at 100. The segments differ a lot in difficulty; the running WER over the full set swung between 0.21 and 0.31. Subsets measure noise.
For reference, a CC BY-NC Swiss German fine-tune of the same base measured 22.8 % at fp16 on the same segments — the unadapted base is better, and sellable.
MIT, from the base model. Copyright (c) 2022 OpenAI. Changes were made: exported to ONNX and quantized to int8 by CHO CHI CHO GMBH, Zürich.