README.md

---
license: cc-by-nc-4.0
language: [en, de]
tags: [automatic-speech-recognition, onnx, onnx-asr, whisper, verbatim-asr]
base_model: nyrahealth/CrisperWhisper
---

# CrisperWhisper — ONNX

ONNX export of [nyrahealth/CrisperWhisper](https://huggingface.co/nyrahealth/CrisperWhisper)
(Whisper large-v3 fine-tuned for verbatim transcription — includes filler words,
disfluencies, pauses) for [onnx-asr](https://github.com/istupakov/onnx-asr) (standard
`whisper` model type — works with stock onnx-asr, no patches needed). Addresses
[onnx-asr#140](https://github.com/istupakov/onnx-asr/issues/140).

fp32 and int8 variants included; int8 decoder produced by quantizing the pre-merge
decoders and re-merging (the merged graph's If-subgraphs defeat direct quantization).

**License**: original model is CC-BY-NC-4.0 (nyra health / nyralabs). This ONNX export
is a derivative, non-commercial use only, and inherits the same license and
attribution requirement. Credit: [nyrahealth/CrisperWhisper](https://huggingface.co/nyrahealth/CrisperWhisper).

**Export gotcha (fixed here)**: `optimum`'s ONNX export re-saves the tokenizer and
silently truncates CrisperWhisper's `vocab.json` (698KB vs the source repo's 878KB),
shifting every special-token ID (e.g. `<|startoftranscript|>` came out as 45073
instead of the correct 50258) and producing near-blank/looping decode output. This
export replaces `vocab.json`/`added_tokens.json` with the originals copied verbatim
from the source repo.

## Usage

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

Verified on FLEURS English and German clips: fp32 and int8 outputs match each other
and the native `transformers` pipeline output exactly, for both languages. RTF on an
AMD Ryzen 5 7600 (6C/12T CPU, shared/loaded host): ~1.7-3.0 fp32, ~0.8-1.5 int8.