You can deploy this model and start earning money today!

frictional123/
sherpa-onnx-kb-whisper-int8

Name: sherpa-onnx-kb-whisper-int8
Availability: InStock

KBLab/kb-whisper-large

quantized

apache-2.0

Views

No views yet

sherpa-onnx KB-Whisper int8

sherpa-onnx-compatible INT8 ONNX exports of KBLab's Swedish-tuned Whisper models. Same encoder/decoder split + tokens.txt format as csukuangfj/sherpa-onnx-whisper-*.

Files

Variant	Encoder	Decoder	Total
small	kb-whisper-small-encoder.int8.onnx (112 MB)	kb-whisper-small-decoder.int8.onnx (262 MB)	~376 MB
medium	kb-whisper-medium-encoder.int8.onnx (374 MB)	kb-whisper-medium-decoder.int8.onnx (571 MB)	~947 MB
large	kb-whisper-large-encoder.int8.onnx (767 MB)	kb-whisper-large-decoder.int8.onnx (1.01 GB)	~1.78 GB

Each variant also has a kb-whisper-{size}-tokens.txt (~800 KB, identical across variants — multilingual Whisper tokenizer).

How they were exported

PyTorch safetensors → OpenAI Whisper shell (state dict remapped) → sherpa-onnx encoder/decoder ONNX → INT8 dynamic quantization (MatMul ops only). Export script: see scribe/scripts/kb-whisper-export (private until cleaned up for public release).

Usage with sherpa-onnx (Python)

python

import sherpa_onnx
recognizer = sherpa_onnx.OfflineRecognizer.from_whisper(
    encoder="kb-whisper-large-encoder.int8.onnx",
    decoder="kb-whisper-large-decoder.int8.onnx",
    tokens="kb-whisper-large-tokens.txt",
    language="sv",
    task="transcribe",
    provider="cpu",
)

License

Apache-2.0, inherited from the base KBLab/kb-whisper-* models.

frictional123/sherpa-onnx-kb-whisper-int8