Multilingual Dictation App based on OpenAI Whisper

Multilingual dictation app based on the powerful OpenAI Whisper ASR model(s) to provide accurate and efficient speech-to-text conversion in any application. The app runs in the background and is triggered through a keyboard shortcut. It is also entirely offline, so no data will be shared. It allows users to set up their own keyboard combinations and choose from different Whisper models, and languages.

Prerequisites

The PortAudio and llvm library is required for this app to work. You can install it on macOS using the following command:

brew install portaudio llvm

Permissions

The app requires accessibility permissions to register global hotkeys and permission to access your microphone for speech recognition.

Installation

Clone the repository:

git clone https://github.com/foges/whisper-dictation.git
cd whisper-dictation

If you use poetry:

poetry install
poetry shell

Or, if you don't use poetry, first create a virtual environment:

python3 -m venv venv
source venv/bin/activate

Install the required packages:

pip install -r requirements.txt

Usage

Run the application:

python whisper-dictation.py

By default, the app uses the "base" Whisper ASR model and the key combination to toggle dictation is cmd+option on macOS and ctrl+alt on other platforms. You can change the model and the key combination using command-line arguments. Note that models other than tiny and base can be slow to transcribe and are not recommended unless you're using a powerful computer, ideally one with a CUDA-enabled GPU. For example:

python whisper-dictation.py -m large -k cmd_r+shift -l en

The models are multilingual, and you can specify a two-letter language code (e.g., "no" for Norwegian) with the -l or --language option. Specifying the language can improve recognition accuracy, especially for smaller model sizes.

Replace macOS default dictation trigger key

You can use this app to replace macOS built-in dictation. Trigger to begin recording with a double click of Right Command key and stop recording with a single click of Right Command key.

python whisper-dictation.py -m large --k_double_cmd -l en

To use this trigger, go to System Settings -> Keyboard, disable Dictation. If you double click Right Command key on any text field, macOS will ask whether you want to enable Dictation, so select Don't Ask Again.

Setting the App as a Startup Item

To have the app run automatically when your computer starts, follow these steps:

1. Open System Preferences.
2. Go to Users & Groups.
3. Click on your username, then select the Login Items tab.
4. Click the + button and add the run.sh script from the whisper-dictation folder.