WhisperWriter is a small speech-to-text app that uses OpenAI's Whisper model to auto-transcribe recordings from a user's microphone.
Once started, the script runs in the background and waits for a keyboard shortcut to be pressed (ctrl+shift+space
by default). When the shortcut is pressed, the app starts recording from your microphone. There are three options to stop recording:
voice_activity_detection
that stops recording once it detects a long enough pause in your speech.press_to_toggle
that stops recording when the activation key is pressed again.hold_to_record
that stops recording when the activation key is released.
You can change the activation key and recording mode in the Configuration Options. While recording and transcribing, a small status window is displayed that shows the current stage of the process (but this can be turned off). Once the transcription is complete, the transcribed text will be automatically written to the active window.
The transcription can either be done locally through the faster-whisper Python package or through a request to OpenAI's API. By default, the app will use a local model, but you can change this in the Configuration Options. If you choose to use the API, you will need to provide your OpenAI API key in a .env
file.
Fun fact: Almost the entirety of the initial release of the project was pair-programmed with ChatGPT-4 and GitHub Copilot using VS Code. Practically every line, including most of this README, was written by AI. After the initial prototype was finished, WhisperWriter was used to write a lot of the prompts as well!
Before you can run this app, you'll need to have the following software installed:
- Git: https://git-scm.com/downloads
- Python
3.11
: https://www.python.org/downloads/
If you want to run faster-whisper
on your GPU, you'll also need to install the following NVIDIA libraries:
To set up and run the project, follow these steps:
git clone https://github.com/savbell/whisper-writer
cd whisper-writer
python -m venv venv
# For Linux and macOS:
source venv/bin/activate
# For Windows:
venv\Scripts\activate
pip install -r requirements.txt
To switch between running Whisper locally and using the OpenAI API, you need to modify the src\config.json
file:
- If you prefer using the OpenAI API, set
"use_api"
totrue
. You will also need to set up your OpenAI API key in the next step. - If you prefer using a local Whisper model, set
"use_api"
tofalse
. You may also want to change the device that the model uses; see the Model Options. Note that you need to have the NVIDIA libraries installed to run the model on your GPU.
{
"use_api": false, // Change this value to true to use the OpenAI API
...
}
Copy the ".env.example" file to a new file named ".env":
# For Linux and macOS
cp .env.example .env
# For Windows
copy .env.example .env
Open the ".env" file and add in your OpenAI API key:
OPENAI_API_KEY=<your_openai_key_here>
You can find your API key on the OpenAI dashboard. You will need to have available credits to use the API.
python run.py
WhisperWriter uses a configuration file to customize its behaviour. To set up the configuration, modify the src\config.json
file:
{
"use_api": false,
"api_options": {
"model": "whisper-1",
"language": null,
"temperature": 0.0,
"initial_prompt": null
},
"local_model_options": {
"model": "base",
"device": "auto",
"compute_type": "auto",
"language": null,
"temperature": 0.0,
"initial_prompt": null,
"condition_on_previous_text": true,
"vad_filter": false
},
"activation_key": "ctrl+shift+space",
"recording_mode": "voice_activity",
"sound_device": null,
"sample_rate": 16000,
"silence_duration": 900,
"writing_key_press_delay": 0.005,
"remove_trailing_period": false,
"add_trailing_space": true,
"remove_capitalization": false,
"print_to_terminal": true
}
use_api
: Set totrue
to use the OpenAI API for transcription. Set tofalse
to use a local Whisper model. (Default:false
)api_options
: Contains options for the OpenAI API. See the API reference for more details.model
: The model to use for transcription. Currently onlywhisper-1
is available. (Default:"whisper-1"
)language
: The language code for the transcription in ISO-639-1 format. (Default:null
)temperature
: Controls the randomness of the transcription output. Lower values (e.g., 0.0) make the output more focused and deterministic. (Default:0.0
)initial_prompt
: A string used as an initial prompt to condition the transcription. Here's some info on how it works. Set to null for no initial prompt. (Default:null
)
local_model_options
: Contains options for the local Whisper model. See the function definition for more details.model
: The model to use for transcription. See available models and languages. (Default:"base"
)device
: The device to run the local Whisper model on. Options includecuda
for NVIDIA GPUs,cpu
for CPU-only processing, orauto
to let the system automatically choose the best available device. (Default:auto
)compute_type
: The compute type to use for the local Whisper model. More information can be found here. (Default:auto
)language
: The language code for the transcription in ISO-639-1 format. (Default:null
)temperature
: Controls the randomness of the transcription output. Lower values (e.g., 0.0) make the output more focused and deterministic. (Default:0.0
)initial_prompt
: A string used as an initial prompt to condition the transcription. Here's some info on how it works. Set to null for no initial prompt. (Default:null
)condition_on_previous_text
: Set totrue
to use the previously transcribed text as a prompt for the next transcription request. (Default:true
)vad_filter
: Set totrue
to use a voice activity detection (VAD) filter to remove silence from the recording. (Default:false
)
activation_key
: The keyboard shortcut to activate the recording and transcribing process. (Default:"ctrl+shift+space"
)recording_mode
: The recording mode to use. Options includevoice_activity_detection
to use voice activity detection to determine when to stop recording, orpress_to_toggle
to start and stop recording by pressing the activation key, orhold_to_record
to start recording when the activation key is pressed down and stop recording when the activation key is released. (Default:"voice_activity"
)sound_device
: The name of the sound device to use for recording. Set tonull
to let the system automatically choose the default device. To find a device number, runpython -m sounddevice
. (Default:null
)sample_rate
: The sample rate in Hz to use for recording. (Default:16000
)silence_duration
: The duration in milliseconds to wait for silence before stopping the recording. (Default:900
)writing_key_press_delay
: The delay in seconds between each key press when writing the transcribed text. (Default:0.005
)remove_trailing_period
: Set totrue
to remove the trailing period from the transcribed text. (Default:false
)add_trailing_space
: Set totrue
to add a trailing space to the transcribed text. (Default:true
)remove_capitalization
: Set totrue
to convert the transcribed text to lowercase. (Default:false
)print_to_terminal
: Set totrue
to print the script status and transcribed text to the terminal. (Default:true
)hide_window
: Set totrue
to hide the status window.
If any of the configuration options are invalid or not provided, the program will use the default values.
You can see all reported issues and their current status in our Issue Tracker. If you encounter a problem, please open a new issue with a detailed description and reproduction steps, if possible.
Below are features I am planning to add in the near future:
- Restructuring configuration options to reduce redundancy
- Update to use the latest version of the OpenAI API
- Additional post-processing options:
- Simple word replacement (e.g. "gonna" -> "going to" or "smiley face" -> "π")
- Using GPT for instructional post-processing
- Updating GUI
- Creating standalone executable file
Below are features not currently planned:
- Pipelining audio files
Implemented features can be found in the CHANGELOG.
Contributions are welcome! I created this project for my own personal use and didn't expect it to get much attention, so I haven't put much effort into testing or making it easy for others to contribute. If you have ideas or suggestions, feel free to open a pull request or create a new issue. I'll do my best to review and respond as time allows.
- OpenAI for creating the Whisper model and providing the API. Plus ChatGPT, which was used to write a lot of the initial code for this project.
- Guillaume Klein for creating the faster-whisper Python package.
- All of our contributors!
This project is licensed under the GNU General Public License. See the LICENSE file for details.