The README gets you running. This page covers every setting in more detail.
- The window
- Hotkeys
- Audio
- Speech engines
- Translation and suggestion models
- About you
- Configuration file
- Orange-outlined lines are what the other person said: their words on top, the
translation underneath. The badge in the corner (
Ctrl+Alt+1,Ctrl+Alt+2, …) is the hotkey that asks for replies to that sentence:1is the newest,2the one before, and so on. - "You" lines are what you said, when your microphone is on.
- The bottom panel shows reply suggestions in their language, with your language underneath. Click one to copy it.
- The lock button (🔒, next to Suggest) keeps the view on the newest sentence.
Scrolling up by hand unlocks it (🔓) so the view stays where you are reading; click the
lock or press
Ctrl+Alt+Lto follow again. - The language label in the title bar (for example "French (locked)") is the language Polyglot settled on. Click it to let it detect the language again.
These work system-wide, even while another app has focus.
| Action | Default |
|---|---|
| Suggest replies to the last sentence | Ctrl+Alt+S |
| Suggest replies to one of their recent sentences | Ctrl+Alt+1 (newest) … Ctrl+Alt+4 |
| Show / hide the window | Ctrl+Alt+H |
| Lock / unlock following the latest sentence | Ctrl+Alt+L |
Inside the window: Ctrl+Enter suggests, Esc hides.
Settings → Window sets how many recent sentences get a numbered hotkey (1 to 10, following
the number row: 1–9, then 0) and which keys are held with the number.
On macOS, global hotkeys need permission: System Settings → Privacy & Security → Accessibility (and Input Monitoring), then turn on Polyglot.
- Windows: Polyglot hears whatever your computer is playing (the call, a video), with no extra setup. Technically this is WASAPI loopback capture.
- macOS: macOS does not let apps hear other apps' sound directly, so you need a free
virtual audio device:
- Install BlackHole 2ch.
- In Audio MIDI Setup, create a Multi-Output Device that includes both your speakers or headphones and BlackHole, and set it as your output. You still hear the call, and BlackHole gets a copy.
- In Polyglot, Settings → Audio → Listen to → BlackHole.
From a source install, uv run polyglot --list-devices lists every input and loopback
device.
Settings → Audio → "Also listen to my microphone" adds a second input for your own voice (off by default). What you say is transcribed on this computer, shown as a "You" line, and used as context, so suggestions know what you already said. It is never translated and never replied to.
Use headphones: with speakers, the microphone also hears the other person and those lines would be tagged as yours. If the microphone cannot be opened you get a warning and listening to the other person carries on.
The Test button next to the microphone picker records 5 seconds and plays them back, then tells you if it was too quiet, too loud, or silent. Stop listening first: the playback would otherwise be picked up as the other person's audio.
Settings → Speech → Speech engine picks the speech recognizer. All of them run on your computer, and none needs an NVIDIA graphics card.
| Engine | In the download? | Languages | Notes |
|---|---|---|---|
| Whisper (default) | yes | 99 | Good all-rounder. Bigger models (Settings → Speech → Whisper model) are more accurate and slower. base is about 150 MB. |
| Parakeet v3 | yes | 25 European, including French | Fastest and most accurate on a CPU. About 670 MB. It does not detect the language, so set "Their language". |
| Qwen3-ASR | no, source install only | 30, plus Chinese dialects | Detects the language. 0.6B suits a CPU; 1.7B wants a GPU. Needs PyTorch (about 1 GB). |
Each engine downloads its model the first time you use it.
From source, install the optional engines with uv sync --extra parakeet,
uv sync --extra qwen3, or both: uv sync --extra parakeet --extra qwen3.
Run on: Automatic uses a graphics card when the engine can use one, and the CPU
otherwise. For Parakeet on an AMD or Intel graphics card (source install), replace ONNX
Runtime with its DirectML build: uv pip uninstall onnxruntime && uv pip install onnxruntime-directml (redo it after each uv sync).
Settings → Translation and Settings → Suggestions each have a model dropdown. It starts
with a short built-in list and, once an OpenAI key is set, loads the chat models your key
can use (Refresh reloads it). Any model ID can also be typed in. The default for both is
gpt-6-luna; a saved setting that still names a retired model (gpt-4o-mini,
gpt-4.1-mini) is moved to it when the config is loaded.
Translation and suggestions can use different models: a larger model for suggestions
(for example gpt-6.1-sol) leaves translation fast.
How requests are made
Requests reuse one HTTPS connection per worker thread, so only the first one pays for the
connection setup. Older models get a temperature; reasoning models get the lowest
reasoning_effort that fits the job instead (none for translation, low for
suggestions). If a model rejects a parameter or value, the request is retried once without
it, and that is remembered for the session.
Settings → Translation stores keys in your system's keychain (Windows Credential Manager
or macOS Keychain), never in a file. For development, POLYGLOT_OPENAI_API_KEY /
OPENAI_API_KEY and POLYGLOT_GOOGLE_API_KEY environment variables override the keychain.
Settings → About you lets suggestions talk about you: your name, role, skills, languages,
experience, hobbies and interests, how you want replies to sound, and what you want from
the conversation. Fill it in once; it is saved in config.toml (the [profile] section)
and used every time, with no restart.
Everything in your profile is sent to the AI provider with each suggestion request, so it only has fields for things you would say out loud to the other person. There is deliberately no field for health, politics, religion, money or other private details, and the free-text fields are yours to keep clean.
Settings → Suggestions → Extra instructions adds your own rules to every request (up to 1000 characters), for example "Don't ask questions back unless the conversation needs one."
Suggestions also see the conversation around the sentence they answer: up to
suggestions.history lines before it and, when a numbered hotkey answers an earlier
sentence, up to suggestions.history_after after it, within suggestions.history_chars
characters in all (the nearest are kept first).
Settings are saved to config.toml in your user config folder:
- Windows:
%LOCALAPPDATA%\polyglot - macOS:
~/Library/Application Support/polyglot
config.example.toml lists every option with its default.
Precedence, lowest to highest: defaults, then config.toml, then
POLYGLOT_<SECTION>__<FIELD> environment variables (e.g. POLYGLOT_ASR__MODEL=small),
then changes made in the settings dialog.
Logs are written to the user log folder and contain no transcript text unless
privacy.log_content is turned on.