README на русском

Real-time speech translator for video/voice calls. Translates both sides of the conversation live — you speak your language, the other person hears theirs, and vice versa.

How it works: Your mic audio goes through Speech-to-Text, gets translated by an LLM, then synthesized back to speech and routed into your call. The same happens in reverse for the other person's audio.

Supports 29 languages with STT, translation, and TTS. Voice models from Piper — download any language directly from the web UI.

macOS License GitHub stars

Note: macOS production today (14+). Windows port in active development (docs/windows.md). Linux port started on feat/linux — contributions welcome.

git clone https://github.com/org-event/OpenPolySphere.git
cd OpenPolySphere
./scripts/bootstrap # dev deps + git hooks (like bun install)
cargo run --release -p translator -- setup # download models (first time)
cargo run --release -p translator # start server

Open http://127.0.0.1:5050 in Google Chrome.

Local mode (default): Whisper STT + Opus-MT translation — no API keys required. Cloud STT/translation optional via Settings.

./scripts/bootstrap is the first command after git clone — like bun install in a JS project. It installs the just task runner if needed, then runs just install.

What just install does:

If just is already installed, you can run just install directly instead of ./scripts/bootstrap.

Common commands (run from the repo root):

After install: just setup once, then just run (or cargo run --release -p translator). Optional: cp .env.example .env for cloud API keys.

Single Rust binary (translator): Axum web server on :5050 + in-process audio engine (STT, translation, TTS).

Browser (app.js) ←SSE→ Axum ←→ audio-core Engine ←→ CoreAudio / models

Optional API keys (cloud STT/translation): Deepgram, OpenRouter

If you prefer step-by-step setup:

xcode-select --install
brew install rustup espeak-ng onnxruntime
rustup-init -y --default-toolchain stable
source ~/.cargo/env

Download and install from existential.audio/blackhole.

You need both:

• BlackHole 16ch — captures audio from your call app (Google Meet, Zoom, etc.)
• BlackHole 2ch — sends translated audio back to the call

Setup in your call app (Google Meet, Zoom, etc.):

1. Open the call in Google Chrome (not Safari)
2. Set BlackHole 2ch as the microphone in the call app
3. Set BlackHole 16ch as the speakers in the call app

Note: Do NOT use a Multi-Output Device — it may cause audio issues. Set BlackHole devices directly in the call app settings.

TTS voices come from Piper. Run cargo run --release -p translator -- setup to download default voices, Whisper, and Opus-MT models. Additional voices can be downloaded from the web UI.

To download manually:

mkdir -p models/piper-en models/piper-ru
# English (default)
curl -sL https://huggingface.co/rhasspy/piper-voices/resolve/v1.0.0/en/en_US/ryan/medium/en_US-ryan-medium.onnx \
 -o models/piper-en/en_US-ryan-medium.onnx
curl -sL https://huggingface.co/rhasspy/piper-voices/resolve/v1.0.0/en/en_US/ryan/medium/en_US-ryan-medium.onnx.json \
 -o models/piper-en/en_US-ryan-medium.onnx.json
# Russian (default)
curl -sL https://huggingface.co/rhasspy/piper-voices/resolve/v1.0.0/ru/ru_RU/denis/medium/ru_RU-denis-medium.onnx \
 -o models/piper-ru/ru_RU-denis-medium.onnx
curl -sL https://huggingface.co/rhasspy/piper-voices/resolve/v1.0.0/ru/ru_RU/denis/medium/ru_RU-denis-medium.onnx.json \
 -o models/piper-ru/ru_RU-denis-medium.onnx.json

Browse all available voices at rhasspy.github.io/piper-samples.

cp .env.example .env

Edit .env:

DEEPGRAM_API_KEY=your_key_here
GROQ_API_KEY=your_key_here
ORT_DYLIB_PATH=/opt/homebrew/lib/libonnxruntime.dylib

cargo run --release -p translator -- setup # first time: download models
cargo run --release -p translator # start server

Open http://127.0.0.1:5050 in Chrome.

• Live transcript — chat-style bubbles with original text and translation
• 29 languages — switch language pair from Settings, download voices with one click
• Voice selection — multiple voices per language with preview playback
• Audio monitor — hear translations in your browser (Chrome only)
• Start/Stop — control the engine without restarting
• Mute — independently mute outgoing or incoming pipelines
• Bookmarks — star important phrases, filter to show only starred
• Export — download the full transcript as a text file
• Compact/Full view — toggle between detailed and compact transcript
• Latency metrics — per-phrase STT, translation, TTS, and total latency
• Dark/Light theme — toggle with persistence

TTS requires downloading a Piper voice model for the language (one-click from the web UI). Japanese and Korean have STT and translation but no Piper TTS voice available.

"Engine not starting"

• Press Start after the page loads (server runs idle until then)
• For local mode: models in models/ — run cargo run --release -p translator -- setup
• Verify ORT_DYLIB_PATH points to your onnxruntime library
• Run cargo build -p translator to check for build errors

"No audio from call"

• Ensure BlackHole 16ch is set up in a Multi-Output Device
• Check that your call app uses BlackHole 2ch as its microphone

"TTS not working"

• Verify espeak-ng is installed: espeak-ng --version
• Check that voice model files exist in models/piper-{lang}/
• Download voices from Settings in the web UI

"No sound in monitor"

• Use Chrome — Safari does not support audio output routing required for monitor
• Check your system audio output is set to speakers (not BlackHole)

"OpenRouter key shows invalid"

• Only needed when cloud translation is enabled
• Keys in .env work even if the Settings field is empty

MIT — see LICENSE. Copyright (c) 2026 Kai Letov (original author).