Speak listens when you tap a hotkey, transcribes your voice with Whisper, cleans up the "um"s and false starts with a small local language model (Qwen 2.5), and types the result into the app you are using: Word, Outlook, Teams, a browser, a terminal, anything.

Everything runs on your own CPU. No audio or text leaves your computer, there is no account, and you do not need administrator rights.

---

Features

• 100% offline and private. Audio and transcripts stay in your computer's memory. Nothing is sent to a server.
• Works in any app. Text is pasted or typed into whichever window has focus. Terminals and legacy apps that reject Ctrl+V get direct typing instead.
• Cleans up how people actually talk. Removes fillers (um, uh, øh, altså, äh, euh), fixes stutters and false starts, and turns spoken commands like "bullet point" and "new line" into real formatting.
• Multilingual. English, Danish, German and French, auto-detected or locked to one language.
• Learns your words. Add names, products and jargon to a personal dictionary. If Speak mishears a word, fix it in your document, select the sentence, press Ctrl+Alt+L, and Speak remembers the correction.
• No admin rights. Runs entirely in Windows user space, with no services, drivers or firewall prompts.
• Fast on a laptop CPU. Short phrases appear in about 1 second, and longer dictations in about 2 to 3 seconds.

How it works

flowchart LR
    A[Tap hotkey<br/>and speak] --> B[Whisper<br/>speech to text]
    B --> C[Personal dictionary<br/>fixes known mishearings]
    C --> D[Qwen 2.5 on CPU<br/>removes fillers, formats]
    D --> E[Text typed into<br/>the active app]

Short, clean phrases such as "Yes, that sounds good." skip the language model entirely, so they appear almost instantly.

Screenshots

Install

The easy way (no technical knowledge needed)

1. Press the Start button, type PowerShell, and open it.

2. Copy this line, paste it into the blue window (right-click pastes), and press Enter:

   irm https://raw.githubusercontent.com/xav-art/Speak/main/install.ps1 | iex
   

3. Wait 5 to 15 minutes. The installer explains each step, and Speak starts by itself when it is done.

From then on, start Speak from the Speak icon on your Desktop or in the Start Menu. To update, run the same line again: your settings, dictionary and models are kept.

What the installer does, all inside your user account and without admin rights:

• Installs Python 3.12 for your user if you do not have Python 3.10 to 3.13 (with winget, or the signed installer from python.org).
• Puts Speak in %LOCALAPPDATA%\Speak with its own private Python environment, so it does not change anything else on your PC.
• Downloads the speech model and the AI model once (about 1.3 GB) and checks the AI model's SHA-256 checksum.
• Adds Speak and Uninstall Speak to the Start Menu, a Desktop icon, and the speak command for terminals.

No .exe is downloaded or run apart from Python's own signed installer, so antivirus and SmartScreen have nothing to flag. You can read install.ps1 before you run it.

From a ZIP file

1. Click the green Code button on this page and choose Download ZIP.
2. Extract the ZIP to the folder where you want Speak to live, for example Documents\Speak.
3. Double-click setup.cmd. If Windows asks whether to run a downloaded file, click Run: this happens only once, because setup removes the download mark from Speak's files.

Uninstall

Open the Start Menu and choose Uninstall Speak. It asks whether to keep your personal dictionary and settings.

Requirements

• Windows 10 or 11 (64-bit)
• A microphone
• About 2 GB of free disk space and 4 GB of RAM or more
• Internet access during setup only

For developers

git clone https://github.com/xav-art/Speak.git
cd Speak
python -m venv .venv
.venv\Scripts\python -m pip install --only-binary :all: --no-deps --index-url https://abetlen.github.io/llama-cpp-python/whl/cpu llama-cpp-python
.venv\Scripts\python -m pip install -r requirements.txt
.venv\Scripts\python -X utf8 speak.py

llama-cpp-python comes from its project's prebuilt CPU wheel index because PyPI only has the source, which needs a C++ compiler. Missing models download automatically on first start. You can also run setup.cmd in your clone to do all of this in place.

Usage

Start Speak from its Desktop or Start Menu icon, or double-click start_speak.bat. Keep its window open while you work; you can minimize it.

1. Click where you want to write: an email, a document, a chat box.
2. Tap < (the key next to Left Shift) and speak.
3. Tap < again. The cleaned text appears at your cursor.

Keyboard controls

Type these in the Speak window:

Command-line options: --hotkey, --model, --language, --paste-mode, --ai-mode. Run speak --help for details.

Personal dictionary

Speak keeps your word list in speak_dictionary.json, next to speak.py. It never leaves your PC.

What Speak does with your words:

1. Whisper is primed with your dictionary before every dictation, so it hears the right spelling in the first place.
2. Learned mishearings are corrected automatically once they are confirmed: when the same context appears again, or after three corrections. Until then the AI cleanup gets a one-line hint.
3. Exact casing is enforced for multi-word and unusually cased entries (øresund bridge becomes Øresund Bridge, sap becomes SAP).

Auto-learn ignores rewrites (more than half the words changed), everyday words, and anything shorter than three letters.

Configuration

Settings are saved to speak_config.json when you change them in the app. You can also edit the file directly:

Performance notes

Measured on an Intel Core Ultra 7 265U with faster-whisper-base and Qwen 2.5 1.5B:

Tested and rejected (kept as opt-in config keys): prompt-lookup speculative decoding (speculative_decoding, 2.5x slower on CPU with llama-cpp-python 0.3.x), flash attention (no gain), and Whisper hotwords (makes multi-word names stutter). On hybrid P/E-core CPUs, more Whisper threads were consistently slower.

Project structure

Speak/
├── speak.py                  # The app: recording, transcription, AI cleanup, text insertion
├── speak_dictionary.py       # Personal dictionary and learning from corrections
├── test_speak.py             # Unit tests
├── requirements.txt          # Python dependencies
├── install.ps1               # Installer, updater and uninstaller (no admin rights)
├── setup.cmd                 # Double-click to run install.ps1 in this folder
├── uninstall.cmd             # Removes Speak (also in the Start Menu)
├── start_speak.bat           # Double-click launcher (the Desktop icon points here)
├── speak, speak.bat, speak.cmd  # Terminal launchers ("speak" on your PATH)
├── install_dependencies.bat  # Old name for setup.cmd, kept for existing users
├── assets/                   # App icon
├── docs/screenshots/         # Images used in this README
└── models/                   # Downloaded models (not in git)

speak_config.json, speak_dictionary.json, .venv/ and models/ are personal, generated or large, so they are ignored by git.

Running the tests

.venv\Scripts\python test_speak.py

Privacy

Speak makes network requests only during setup and to download a model from Hugging Face the first time you use it. Recording, transcription, cleanup and your dictionary stay on your machine. The microphone is open only between your two hotkey taps.

License

MIT © 2026 xav-art