Views
No views yet
whisper model type — works with
stock onnx-asr, no patches needed). fp32 and int8 variants included.1import onnx_asr
2model = onnx_asr.load_model("whisper", "path/to/this/repo") # or quantization="int8"
3print(model.recognize("audio_16khz.wav", language="id"))id_id) test clip: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).