ssciwr / vink

A stand-alone application with GUI for OpenAI's Whisper

Vink - an OpenAI Whisper standalone distribution

This is a stand-alone application that packages OpenAI's Whisper into a distribution that does not require users to have Python installed. Additionally, it provides a minimalistic graphical user interface for transcription.

Installation on Windows 10+

Please download the zipped distribution from Zenodo and unpack it to a target location of your choice:

• 2024-01-08 Build from zenodo.org, MD5 checksum: d94d38f8db81ea1b9338e152496cc270

Usage on Windows

Please install by double clicking on the downloaded installer. Running vink.exe (e.g. from your start-up menu) will fire up the graphical user interface, allowing you to specify input and output files, a model and device to run on.

If you want to work with whisper's command line interface instead, you can do so by providing arguments on the command_line:

vink --help

Linux

Currently, there is no (working) stand-alone packaging for Linux due to issues in pyinstaller's handling of PySide6. If you are still interested in using the GUI from Ubuntu/Debian, the following sequence of commands will install dependencies and run locally:

sudo apt install ffmpeg python3 python3-pip git
git clone https://github.com/ssciwr/vink.git
cd vink
python -m pip install -r requirements.txt
python vink.py

Licensing

The code provided in this project itself is covered by the MIT license. The overall distribution is also licensed under the MIT license. For details about bundled components and their license, please check LICENSE.md.

vink's Issues

Progress bar in GUI

Ship GPL license text

Required for full legal text of LGPL

Show output directly in GUI and allow copying to clipboard

Make GPU the default in GUI if one was found

Non-Blocking execution of transcription

Suppress terminal popup on Windows

Download progressbar in GUI

Currently, if a model is downloaded on-demand, the progress bar is shown in the terminal...

Cuda device selection from gui

Update to latest released whisper version

Latest release: https://github.com/openai/whisper/releases/tag/v20230918

And double check that everything is still working as expected.

Silence terminal warnings

There are currently several warnings emitted by the CUDA runtime that might confuse users.

updating to the letest version of whisper

there is any way that you can share exact command to use pyinstaller how to compile this package that the command line will still working.
if you interest I have personal project to complete and I'm planning to use whisper as a asr model. i can pay for it
please contact [email protected] if this is something that you interest.

Add Github Pages with description and download links

lightweight c++/cpu-only whisper port

@dokempf I just found this which might be relevant: https://github.com/ggerganov/whisper.cpp

Distribute bundle as a setup wizard

See here and adapt: https://nsis.sourceforge.io/Simple_tutorials

Encoding issues for text output

Transcribing e.g. russian results in encoding issues on Windows. A fix should be to switch to explicitly writing as utf-8.

Pin whisper version

Whisper recently added a release tag. We should pin that tag for robustness.

Allow cancellation of transcription

Follow up to #13

Ship at least one model

To allow offline first use.

Switch to Conda environment instead of pip requirements

To hopefully ship threading+CUDA runtimes for Windows.

Linux script can't find run_whisper.py

Following install instructions for linux:

sudo apt install ffmpeg python3 python3-pip git
git clone https://github.com/ssciwr/whisper-standalone.git
cd whisper-standalone
python -m pip install -r requirements.txt
python run_whisper.py

[Errno 2] No such file or directory

All works well until the last line, can't find the file. Which is not surprising since the file is absent from the repo.

Add GUI based on PySide

Finalize LICENSE.md with bundled system dependencies

Avoid double inclusion of CUDA libraries

Currently, the Linux version ships each CUDA runtime library twice. I think this happened after adding adding the extra index to the pip call.