Dictly Whisper

Open-source offline Whisper dictation app for Windows desktop voice typing and local speech-to-text.

Dictly is an open-source offline Whisper dictation app for Windows built around Whisper, faster-whisper, WhisperX, and local speech-to-text workflows. Hold a hotkey, speak, release, and Dictly transcribes your voice locally and pastes the text into the active app. No API keys, no cloud dependency, no subscription.

What Dictly Is

Dictly is a local AI dictation tool for:

• offline voice typing on Windows
• Whisper speech recognition
• faster-whisper transcription
• push-to-talk speech to text
• desktop dictation in any app

If someone is searching for a Whisper dictation app, offline speech-to-text for Windows, open-source voice typing software, or a faster-whisper desktop transcription tool, this repository is for that use case.

Open Source

Dictly is released under the MIT license in LICENSE.

This repository is the source code. End users do not need to install Python packages if you publish the packaged Windows build from GitHub Releases.

Features

• Offline transcription with Whisper and faster-whisper
• Push-to-talk global hotkey
• Desktop waveform overlay
• Settings UI for backend, model, beam size, language, theme, and device
• Transcription history
• Tray icon and startup option
• CPU fallback when GPU backend is unavailable
• Packaged Windows EXE build for non-technical users

Keywords

Whisper, faster-whisper, whisperx, speech to text, offline dictation, Windows dictation, voice typing, local transcription, AI dictation, desktop dictation, push-to-talk transcription, offline voice input, open source Whisper app, offline speech recognition, Windows voice typing, local AI transcription, desktop speech-to-text.

Download

Latest release:

• Releases page: https://github.com/KushalLukhi/dictly-whisper/releases
• Latest release: https://github.com/KushalLukhi/dictly-whisper/releases/latest

GitHub does not run .exe files inside the repository page itself. For Windows users, the correct flow is:

1. Open the repo's Releases page.
2. Download Dictly-windows.zip.
3. Extract the ZIP on your PC.
4. Open the extracted Dictly folder.
5. Run Dictly.exe.

Do not download only the standalone .exe from inside the folder tree. The app needs the bundled _internal files next to it.

No pip install is required for release users.

Run From Source

python -m venv venv
venv\Scripts\activate
python -m pip install -r requirements.txt
python .\main.py

Build The Windows EXE

python -m venv venv
venv\Scripts\activate
python -m pip install -r requirements-build.txt
.\build.bat

Build output:

• Folder build: dist\Dictly\
• Release ZIP: dist\Dictly-windows.zip

Users can run that packaged build directly without installing Python or project dependencies.

Build A Windows Installer

If you want a normal Windows installer instead of opening the packaged EXE folder directly:

1. Build the app package first.
2. Run:

.\build-installer.bat

Installer output:

• installer-dist\Dictly-Setup.exe

Default behavior:

• uses Inno Setup if available
• otherwise falls back to a built-in IExpress installer

The installer puts Dictly under the current user's AppData\Local\Programs\Dictly folder and creates Start Menu shortcuts.

Build An Offline Windows Installer

If you want users to install Dictly without downloading models on first run, bundle models into the app before building the installer.

Recommended flow:

1. Download the models you want to ship:

.\download-models.bat

2. Build the offline installer:

.\build-offline-installer.bat

Notes:

• The build includes everything found under models\
• The installer will contain those bundled model files
• Larger models make the installer much bigger
• For most users, bundling small is the best size versus quality tradeoff

Publishing Releases On GitHub

Recommended release flow:

1. Build the Windows package locally.
2. Open GitHub and create a new Release.
3. Upload Dictly-windows.zip as a release asset.
4. Tell users to download the ZIP from Releases, extract it, and run Dictly.exe.

Current Backend Notes

• auto: prefers the best available supported backend
• insanely-fast-whisper: CUDA-only transformers backend tuned for fast chunked Whisper inference
• faster-whisper: reliable CPU path with optional CUDA device selection
• whisperx: experimental live mode without alignment or diarization
• whisper-directml: optional and guarded; falls back to CPU if unsupported

Offline Model Setup

Dictly can now load faster-whisper models from disk instead of downloading them from the Hugging Face Hub.

To download every Dictly model locally into the repo:

.\download-models.bat

This downloads:

• tiny
• base
• small
• medium
• large-v2
• large-v3

The files are stored under models\<model-name> and are automatically picked up by the packaged app.

Option 1: pick a local model folder in Settings.

• Open Settings
• Set Model to match the local model you downloaded, for example small
• Set Local Model Directory to the extracted model folder
• Save and reload

Option 2: bundle models into the packaged app.

Place the model files under this repo before building:

models/
  small/
    config.json
    model.bin
    tokenizer.json
    vocabulary.json
    ...

Then rebuild with:

.\build.bat

The packaged app will automatically look for bundled models under models\<model-name> before trying the Hub.

Insanely Fast Whisper Notes

The insanely-fast-whisper backend is intended for NVIDIA CUDA systems with torch + transformers available.

• It is not used by auto; select it explicitly in Settings.
• It falls back to faster-whisper if CUDA or the required Python deps are unavailable.
• It uses OpenAI Whisper checkpoints from Hugging Face such as openai/whisper-small.
• The shared Local Model Directory field must point to a transformers-compatible Whisper model folder for this backend. The bundled models/ layout produced by download_models.py is for faster-whisper and should not be reused here.

Project Layout

app.py
config_manager.py
dictation_overlay.py
history_manager.py
main.py
recorder.py
settings_window.py
startup_manager.py
transcriber.py
transcription_backends.py
tray.py
typer.py
waveform_widget.py
dictly.spec
build.bat
requirements.txt
requirements-build.txt

Notes For Contributors

• The packaged app still downloads Whisper model weights on first use unless a compatible local or bundled model directory is provided.
• First-run model downloads depend on valid TLS certificates. Packaged Windows builds now try the Windows trust store first, then fall back to certifi; machines behind SSL inspection still need their root CA installed in Windows.
• Global hotkeys, microphone access, and text insertion behavior can vary across operating systems.