whisper-large-v3-turbo-german-sherpa

sherpa-onnx int8 export of primeline/whisper-large-v3-turbo-german (Whisper large-v3-turbo fine-tuned for German by Florian Zimmermeister @primeLine, 809M params, WER 2.628% on the German ASR Data-Mix eval).

Published for the Anti-Vocale community model catalog: import as an external Whisper-family model, no app update required.

Files

File	Notes
encoder.int8.onnx	MatMul int8 quantized encoder
decoder.pc-int8.onnx	per-channel int8 decoder (see below)
tokens.txt	standard multilingual Whisper vocab (51866), identical bytes to the other turbo exports

Why per-channel int8 for the decoder

Per-tensor dynamic int8 quantization of a fine-tuned Whisper decoder causes premature end-of-transcript: long audio decodes to about one phrase per 30 s chunk. Fine-tuning shifts weight distributions and per-tensor scales crush the outlier channels. Per-channel quantization keeps one scale per output channel at the same file size; on the German FLEURS test set the per-channel decoder's greedy output is identical to the fp32 decoder's.

Conversion notes

Re-exported from the fine-tuned PyTorch checkpoint (not from any transformers.js-style ONNX mirror, which are structurally incompatible with sherpa-onnx). The decoder's learned positional embedding table is copied verbatim; decoder exported with the legacy TorchScript ONNX exporter (the dynamo exporter bakes an invalid cross-attention KV reshape for this architecture).

Usage

recognizer = sherpa_onnx.OfflineRecognizer.from_whisper(
    encoder="encoder.int8.onnx",
    decoder="decoder.pc-int8.onnx",
    tokens="tokens.txt",
    language="de",
    task="transcribe",
)

Credits

All model quality credit goes to Florian Zimmermeister (@primeline); this repo only repackages the fine-tune for sherpa-onnx. Upstream eval results:

Dataset	WER
Tuda-De	6.441
common_voice_19_0	3.200
multilingual librispeech	2.070
All (mix)	2.628