This script was cobbled together as a result of https://twitter.com/sanchitgandhi99/status/1656665496463495168 and https://old.reddit.com/r/MachineLearning/comments/14xxg6i/d_what_is_the_most_efficient_version_of_openai/jrsbfps/.
This automatically takes a folder full of video files, extracts the audio, and then uses the Whisper JAX API to transcribe them and output SRT files.
I haven't tested this extensively yet but it seems to perform well enough. For me this is a replacement for using Const-Me/Whisper via Whisperer, as using this API is significantly faster.
Note: To make this file directly executable in Linux by calling main.py, uncomment the first line but leave one # so that it reads #!/usr/bin/env python3.
There is a release EXE available. If FFMPEG is present in path or in the same folder, executing this EXE will automatically transcribe all video files in the same folder.
I have a pre-generated main.spec file in the repo, so you can run pyinstaller main.spec.
To build the spec file yourself, these steps need to be followed:
- Run
pyi-makespec --collect-data=gradio_client --collect-data=gradio --onefile main.py - Add the following to the analysis section of main.spec generated:
...
module_collection_mode={
'gradio': 'py', # Collect gradio package as source .py files
},
)```
3. Run `pyinstaller main.spec`.