Follow a conversation in a language you're still learning. Polyglot is a small window that sits beside your call, writes down what the other person says, translates it, and suggests replies you can say back in their language.
A French conversation. Each of their sentences shows the original and the English translation, your own lines are marked "You", and the bottom panel offers two replies to "et de vous prêts ?", in French with the English underneath. Click a reply to copy it.
- Transcribes on your computer. Speech recognition runs locally. Audio never leaves your machine.
- Translates as they speak, sentence by sentence, into your language.
- Suggests replies in their language when you press a hotkey, with your language underneath so you know what you are saying. You can reply to their latest sentence or to any of their last few.
- Knows who you are. Fill in a short profile once (your name, role, what you want from the conversation) and replies sound like you.
- Keeps your notes to yourself. On Windows, the window is left out of screen sharing, so your translations and profile are not broadcast when you share your screen.
Polyglot works for any conversation where you understand some of the other language, but not all of it: a meeting with colleagues abroad, a call with family, a language exchange, a doctor's appointment or a landlord in a new country.
| Windows 10/11 | Download Polyglot-…-windows-x64.zip from the latest release, unzip it anywhere, and open Polyglot.exe inside the folder. |
| macOS (Apple Silicon) | Download Polyglot-…-macos-arm64.zip from the latest release, unzip it, and drag Polyglot into Applications. |
| Intel Macs, Linux | Install from source. |
"Windows protected your PC" or "Apple could not verify Polyglot"? The app is not code-signed yet, so your computer warns you the first time.
- Windows: click More info → Run anyway.
- macOS: open System Settings → Privacy & Security, scroll down, and click Open Anyway next to Polyglot.
The first time you start listening, Polyglot downloads its speech model (about 150 MB). After that it works offline, apart from the optional translation and suggestions.
- Open Polyglot. A small window appears and stays on top of your other windows.
- Choose what to listen to (Settings ⚙ → Audio → Listen to).
- Windows: leave it on the default. Polyglot hears whatever your computer plays.
- macOS: install the free BlackHole audio driver first and pick it here. The user guide walks you through it.
- Set their language (Settings → Speech → Their language), or leave it on automatic.
- Turn on translation and replies (optional). Paste an OpenAI API key in Settings → Translation, tick "Translate", and on the Suggestions tab tick "Suggest replies".
- Press Start and join your call. Each sentence appears when the speaker pauses.
Press
Ctrl+Alt+Sfor replies, and click one to copy it.
These work even while your call app is in front.
| Action | Keys |
|---|---|
| Suggest replies to their last sentence | Ctrl+Alt+S |
| Suggest replies to an earlier sentence | Ctrl+Alt+1 (newest) … Ctrl+Alt+4 (shown on each sentence) |
| Show / hide the window | Ctrl+Alt+H |
| Keep following the newest sentence | Ctrl+Alt+L |
On macOS, allow Polyglot under System Settings → Privacy & Security → Accessibility for hotkeys to work.
Transcription is free and offline. Translation and reply suggestions use OpenAI (or Google, for translation only), which you pay for by use:
- Create an account at platform.openai.com, add a payment method, and create a key under API keys.
- Paste it into Polyglot under Settings → Translation → OpenAI API key.
Each request is a sentence or two of text, so a conversation typically costs very little; see OpenAI's pricing for current rates. The key is kept in your system's keychain, never in a file.
| What | Where it goes |
|---|---|
| Audio | Nowhere. It is transcribed on your computer and discarded. |
| Transcribed text | Sent to OpenAI or Google only if you turn on translation or suggestions. |
| Your profile ("About you") | Sent with each reply request, so it only asks for things you would say out loud. |
| API keys | Your system's keychain (Windows Credential Manager or macOS Keychain). |
| Logs | Kept on your computer, with no transcript text unless you turn it on. |
Ask before you transcribe. Let the people you are talking with know you are using a transcription and translation tool. Rules on recording conversations differ between countries and workplaces, and following them is up to you.
- Hiding from screen sharing works on Windows only. macOS 15 and later no longer let apps opt out of screen capture, so on a Mac the window shows a "visible in screen share" notice instead.
- A sentence appears once the speaker pauses, not word by word. How quickly depends on your computer and the speech engine.
- Accuracy depends on audio quality. Crosstalk, heavy accents and background noise lead to mistakes, so treat the text as a guide.
- Speech engines differ. Whisper covers 99 languages; Parakeet is faster but covers 25 European ones. See speech engines.
- Nothing appears when they talk. Check Settings → Audio → Listen to. On macOS, make sure the call's sound goes to your Multi-Output Device with BlackHole in it.
- My own words show up as theirs. Use headphones. With speakers, the call audio and your microphone pick each other up.
- Wrong language, or gibberish. Set Settings → Speech → Their language instead of automatic. Click the language label in the title bar to let it detect again.
- It's slow. Choose a smaller Whisper model, or switch to Parakeet for European languages (Settings → Speech).
- No translation or replies. Check the API key, and that "Translate" and "Suggest replies" are ticked.
Everything else, from model choices to every setting, is in the user guide.
Needs Git, Python 3.10+ and uv.
git clone https://github.com/Yalhash/PythonPolyglot.git
cd PythonPolyglot
uv sync # creates .venv with the exact locked versions
uv run polyglot --demo # try the window with synthetic audio: no models, keys or network
uv run polyglot # the real thingOptional speech engines: uv sync --extra parakeet and/or --extra qwen3.
uv run polyglot --file recording.wav --fast # transcribe (and translate) a WAV file
uv run polyglot --headless # listen live, print to the terminal
uv run polyglot --list-devices # list input and loopback devices
uv run polyglot --self-check # check every native dependency loadsuv run pytest # fast suite: no models, devices or network
uv run pytest -m slow # real Whisper tiny + Silero (needs espeak-ng); Parakeet/Qwen3 if installed
uv run ruff check . && uv run ruff format .
uv run mypyDependencies are declared in pyproject.toml and pinned in uv.lock; add one with
uv add <package> (or uv add --dev) and commit both files.
CI (GitHub Actions) runs lint, type checks and tests on Windows, macOS and Linux with Python 3.10 and 3.12, plus a job with real models. On feature branches it also commits lockfile updates and ruff autofixes.
The downloads are built with PyInstaller from packaging/polyglot.spec.
To build one locally (on Windows for Windows, on a Mac for macOS):
uv sync --locked --no-dev --extra parakeet
uv pip install "pyinstaller>=6.11"
uv run --no-sync pyinstaller --noconfirm --clean packaging/polyglot.spec
dist/Polyglot/Polyglot.exe --self-check report.txt # macOS: dist/Polyglot.app/Contents/MacOS/PolyglotTo publish a release, push a version tag. The Release workflow
builds both platforms, runs the packaged app's --self-check, and attaches the zips to a
new GitHub release:
git tag v1.0.0 && git push origin v1.0.0Polyglot is about 8,000 lines of typed Python with about 5,000 lines of tests. Its design:
- Ports and adapters. A pure-Python core (standard library and NumPy only) defines interfaces for every outside dependency: audio source, voice detection, speech recognition, translation, suggestions, keychain, screen-capture shield. Each technology lives in its own adapter, so swapping Whisper for Parakeet is one config value.
- The layering is enforced by a test.
test_architecture.pyparses every module's imports and fails the build if the core imports a third-party library, if an adapter reaches into the UI, or if an OS check appears outside the platform layer. - A real-time pipeline with explicit threading. The audio callback only copies into a preallocated ring buffer. One thread does voice detection and cuts sentences at pauses, exactly one thread runs speech recognition in order, a small pool makes network calls with timeouts, and a single coordinator thread owns all conversation state. Results are placed by sentence ID, never by arrival order, and immutable events cross one bridge to the Qt UI thread.
- Graceful degradation. When the recognizer falls behind, the oldest queued sentence is dropped and marked, rather than letting latency grow. A failed translation leaves the transcript in place, and a microphone that won't open never stops the main capture.
- Testable without hardware. Every interface has a working fake, so the full pipeline,
the presenter and the Qt window are tested without models, audio devices or network.
polyglot --demoruns the real app on the same fakes.
The full design, including layers, threads, endpointing rules, error handling and the reasoning behind each decision, is in docs/ARCHITECTURE.md.

