This repo is a fork of whisper ASR models with openvino backend. Currently, the transcribe functionality of all models but large is supported.

To install, please run the following command with the environment described in the origin repo: https://github.com/openai/whisper.git

pip install git+https://github.com/zhuzilin/whisper-openvino.git

And you can use this modified version of whisper the same as the origin version. For example, to test the performace gain, I transcrible the John Carmack's amazing 92 min talk about rendering at QuakeCon 2013 (you could check the record on youtube) with macbook pro 2019 (Intel(R) Core(TM) i7-9750H CPU @ 2.60GHz) with:

whisper carmack.mp3 --model tiny.en --beam_size 3

And the end-to-end time is shown below:

You can check the transcribed txt in carmack.mp3.txt.

All weights and models include the intermediate ONNX are uploaded to huggingface model hub.