Methodology
How metcha actually works.
A clear, source-cited explanation of the audio pipeline that powers a metcha conversation. We name the providers, the models, and the latency budget — and we say where we make trade-offs, because no translator is perfect.
1. Capture
Audio comes in through the iPhone's microphone (or the microphone on the shared earbud, when one is available). metcha uses Apple's voice-activity detection to mark the start and end of each utterance, so you don't have to tap a button to talk. Each utterance is buffered in memory, never written to disk by default.
2. Speech-to-text (STT)
Free tier — Apple's on-device Speech framework. Recognition runs locally; no audio leaves the device. Coverage depends on the device's downloaded language packs. Accuracy is strong for clear speech and the major regional accents; it weakens in noisy environments.
metcha Plus — optional cloud STT via Deepgram's Nova family. Better word-error-rate on accented speech, noisy rooms, and overlapping voices. Audio is sent over a TLS WebSocket, transcribed in near real time, and never retained by Deepgram for training (per our enterprise contract).
3. Translation
Free tier — Apple's on-device Translation framework. Language packs download once; subsequent translations run locally with no network call. The phrasing is competent but literal; it occasionally misses idioms.
metcha Plus — optional "Better Translation" path through an LLM (currently Anthropic's Claude family). The prompt frames each utterance in conversational context, so the translation reads like something a person would actually say, not a phrasebook. When the network is unhealthy, metcha falls back to the Apple on-device path silently — the conversation never stalls waiting on a cloud round-trip.
4. Text-to-speech (TTS)
Free tier — iOS built-in voices via AVSpeechSynthesizer. Quality depends on whether you've installed the "Enhanced" or "Premium" variant for the target language. metcha respects whatever you have.
metcha Plus — premium voices from ElevenLabs's multilingual catalog. Each supported language has a hand-picked short list of native-speaker voices. Optionally, you can clone your own voice once (with explicit, in-app consent) and have your translated lines spoken in something close to it.
Latency budget
A typical turn — speaker finishes, listener hears the translation — runs:
- Capture (VAD endpointing): ~250 ms
- STT (on-device): ~300–600 ms; STT (Deepgram streaming): ~150–300 ms
- Translation (on-device): ~50–150 ms; translation (Claude): ~600–1100 ms
- TTS (Apple, local): ~200–400 ms; TTS (ElevenLabs streaming): ~400–900 ms
End-to-end: roughly 1.0–1.5 s on the free path and 1.5–2.5 s on metcha Plus with all cloud stages active. Both are inside the threshold where a conversation still feels like a conversation; both are noticeably longer than a face-to-face exchange in a single language. We're transparent about it because we'd rather you set expectations correctly than be surprised.
What stays on device
- The free tier is end-to-end on-device. No audio, transcript, or translation leaves the phone.
- metcha Plus features that pass audio or text to a cloud provider do so under TLS to the provider only. We don't proxy or log content on our own servers.
- Voice cloning, when opted in, stores a single voice fingerprint with ElevenLabs under our account. You can delete it at any time from Settings → metcha Plus → Voice cloning → Delete my voice.
- Per-session transcripts are kept on your phone, encrypted at rest by iOS. You can export or delete them whenever.
What we don't do
- We don't train models on your conversations. None of our providers retain audio for model training under our contracts.
- We don't operate a cross-device account system. There's no metcha login, no metcha social graph.
- We don't claim parity with a human interpreter. metcha is an excellent tool for conversational, daily-life translation. For medical, legal, or court-of-record settings, hire a certified human interpreter.
Where to read more
- How metcha works — the user-facing walk-through.
- Privacy policy — the full data-handling story, provider by provider.
- Supported languages — the per-language matrix of STT, translation, and voice coverage.
Corrections
If anything on this page is wrong or out of date, email hello@metcha.io with the correction. We'll update the page and date the change.