WhisperX Server

This is a simple Server Application that receives audio file paths via the endpoint POST transcribe, it then uses Whisper (WhisperX to Transcribe and allign the transcription and outputs the result (transcript.txt and subtitles.srt files) to the directory of the input file, or to the directory that was indicated in the request. An additional JSON file transcription_metadata.json with some metadata will be added as well.

For the server specification (request structure and response behavior) see the OpenAPI specificaiton in swagger.yaml.

This server will use the Whisper model size large-v2. To improve performance on account of accuracy, change the value of WHISPER_MODEL in run.sh to either of `["large", "large-v2", "medium", "small", "tiny"]

For any other documentation refer to WhisperX readme.

Some Notes:

• In order to not spend resources on loading the model - we load it once (lazily) and reuse it afterwards. As a result - the model type and size is specified via the environment variable WHISPER_MODEL.
• The server will only process one video at a time and reject incoming requests while processing is taking place. We can change this to a queue like behaviour in the future.
• The correct Python version to run this is 3.8. Avoid unpleasantness by sticking to 3.8.
• Note that Python 3.8 should be used to install dependecies (pip with Python 3.8 was used succesfully)
• After installing the pre-requirsites as indicated in the WhisperX repository, run the Server by executing the script run_gpu.sh to execute with CUDA or run_cpu.sh for running on the CPU (slow).
• For convenience sake - here's a curl command to trigger the endopint:

For Transcribing a file in the original language:

curl -X POST http://<machine_IP>:8080/transcribe \
    -H "Content-Type: application/json" \
    -d '{"audioPath": "examples/sample_file.mp4"}'

• In order to translate to English, add the query parameter "task": "translate"

curl -X POST http://<machine_IP>:8080/transcribe \
    -H "Content-Type: application/json" \
    -d '{"audioPath": "examples/sample_file.mp4", "task": "translate"}'