whisperX-REST

The simplest yet fully functional RESTful API for whisperX.

Requirements 📃

• Docker.
• GPU with NVIDIA CUDA support (although you can use a CPU, but it will be painfully slow).
• 100+ GiB disk space.

Download ⤵️

git clone https://github.com/AxesAccess/whisperx-rest.git && cd whisperx-rest

Deploy ⚒️

Create volume to store models.

docker volume create whisperx_volume

Due to dependencies building container will take time and space. You may need to repeat if failed.

docker compose up -d

Use 🖥️

First launch will trigger the download of additional files, which will slow down the result. If you get an error, try setting compute_type parameter to int8.

API returns result in a single output format, so you need to set output_format argument to json, srt, tsv, txt, or vtt instead of default all.

If you choose model other than small (the default), the new model will be downloaded and stored in the container.

Note that large models are demanding on GPU memory. You may want to adjust batch_size and compute_type parameters.

Swagger 🐰

Try Swagger UI in browser at http://127.0.0.1:5001/api/docs/.

Curl 🤓

Send audio in a POST request:

curl -s -X POST "http://127.0.0.1:5001/whisperx?output_format=txt" -F "audio=@sample/CNVSample227.wav"

Pass command line parameters for whisperx in the query string like that:

curl -s -X POST "http://127.0.0.1:5001/whisperx?model=medium&batch_size=4&fp16=False&compute_type=int8&language=en&output_format=json" -F "audio=@sample/CNVSample227.wav"

Finally 🙏

Obviously it's not for production use. Do not expose it to public networks.