wsl-cuda-whisper

The audio.whisper R package allows users to easily use OpenAI's Whisper model (e.g., for automated transcription of audio files) from R. Significant speedups can be achieved on machines with CUDA-enabled graphics cards, but setting this up can be complicated. This docker image allows a user on Windows to easily install all the dependencies needed to run audio.whisper with CUDA support via Windows Subsystem for Linux (WSL2). It is built on top of the rocker/tidyverse image, which means it comes with RStudio Server installed.

Tag	Base Image	Operating System	R ver	CUDA ver
latest	rocker/tidyverse	Ubuntu 24.04 LTS	4.4.2	12.6
novad	rocker/tidyverse	Ubuntu 24.04 LTS	4.4.2	12.6
vad	rocker/tidyverse	Ubuntu 22.04 LTS	4.4.1	11.8

Usage:

1. Verify that your machine's graphics card supports CUDA: https://developer.nvidia.com/cuda-gpus
2. On Windows, install the latest game-ready driver from NVIDIA: https://www.nvidia.com/Download/index.aspx#
3. On Windows, install the latest version of Docker Desktop: https://www.docker.com/products/docker-desktop/
4. Open Docker Desktop and click the Terminal button on the bottom of the screen
5. In the Terminal, type docker pull jmgirard/wsl-cuda-whisper:vad or docker pull jmgirard/wsl-cuda-whisper:novad
6. In the Terminal, type docker run --gpus all -it -e PASSWORD=pass -p 8787:8787 jmgirard/wsl-cuda-whisper
7. Once the Terminal has a line beginning with "TTY detected.", the container is ready
8. In Docker Desktop, click the Containers tab on the left and click the "8787:8787" link
9. Your browser should show a login page, enter "rstudio" as the username and "pass" for the password
10. You should now be shown the RStudio page, so enter library(audio.whisper)
11. Now you can download and load whisper models via, e.g., model <- whisper("tiny", use_gpu = TRUE)
12. You can now use the model object and the predict() function with great speed