Whisperizer - Transcription tool using OpenAI's Whisper

Transcription is a tedious work, often very time-consuming and prone to errors. OpenAI developed a model to make this work more efficient: Whisper. Using this, you can relax and let your machine do all the work. However, you need some programming skills to actually use the API.

As automated transcription is something with a very broad target audience, I created this Graphical User Interface (GUI)-based tool called Whisperizer, which brings along multiple additions such as translation, model selection, multiple file support, automated file naming and others to the transcription realm.

What can it do

Note that this work is in progress. The vision is a (Web)App, available either locally or directly online with access to a CUDA-enabled GPU machine, with multiple additions and Quality of Life improvements for an actual easy-to-use, multi-purpose transcription suite.

So far available

• Transcribe single or multiple audiofiles in common formats to textfiles
• Translate the transcription to a variety of languages
• A GUI for more accessible transcription and translation services

Planned (see issues)

• Diarization (speaker detection)
• GUI enhancement and bugfixing (Cancel button and progress bar)
• Improve file management and enable conversion of input files
• Enhance documentation and setup adaptation (especially CUDA-less environments)
• Create a standalone (Web)App to make this actually accessible, as so far it still needs setup/CLI knowledge

Installation

There are two supported options to create new virtual environments and install the according dependencies, Mamba/Conda and Venv + Pip. Using Mamba is recommended.

Using Mamba (if using Conda, replace mamba with conda below):

mamba env create --file environment.yaml
mamba activate whisperizer

Using venv and pip (built-in):

sudo apt install python3.11 # if not installed already
python3.11 -m venv .venv
source .venv/bin/activate
python3.11 -m pip install -r requirements.txt

Note about required space
Despite the low number of dependencies, the environment will still have a size of around 5.5 GB. Depending on the selected model, another 73 MB (tiny) to 3 GB (large-v1) space will be required. The default model turbo needs 1.6 GB space and delivers the best balance between resource cost and performance.

Usage

Start the GUI with:

python src/main.py

You don't need to do anything else, the rest is done inside the GUI.

Note about expected runtime
If you have a modern CUDA-enabled GPU on your machine (e.g., RTX 3060 Ti), the tool is quite fast (talking mere seconds per minute of audio). However, GPU-less systems using only the CPU will take considerably longer (talking nearly 1:1 processing). The values are just an orientation what to expect, this strongly depends on your machine.

GUI

The following image shows the GUI after a finished transcription and translation with all options checked on an Ubuntu (Linux) system:

By hovering over the elements, you get tooltips with a short description what they do. The main elements are:

1. File management (input file, output folder, output file name)
2. Selecting the model (more info here, also available by clicking the info icon)
3. Selecting the language for translation and the option to keep the transcription in addition to the translation
4. A logging window which display status and current task with a progress bar
5. Control buttons to start and cancel the transcription as well as closing the GUI

Contribution

Any advice, proposal for improvement, or question is welcome on all communication channels!

If you want to contribute code, please use pre-commit by running:

pre-commit install

Using this, it is ensured that the code you provide suits the custom formatting rules before committing or creating a pull request.

License

This project is licensed under the MIT License - see the LICENSE file for details.