How your voice is processed. What stays on device. What leaves. And why it matters.
Every phone on the market is listening. Siri and Google Assistant run continuous wake-word detection — audio segments sent to Apple and Google servers on every trigger. Voice recordings retained for up to six months. Apps with microphone permission can activate audio capture while active with no visible indicator.
Meta is more aggressive. Facebook, Instagram, and WhatsApp have been found to activate the microphone beyond user-initiated actions. Your voice is an advertising signal — you mention a product, an ad appears. This is the business model, not a bug.
Telegotchi's microphone activates in one circumstance only: when you press and hold the physical PTT button. There is no wake-word detection. There is no ambient listening. When you release the button, the microphone stops.
This is a hardware guarantee, not a software promise.
Your voice is processed by Whisper running on a Mac Mini you own, on your local network. Audio never reaches the internet. Only the text transcript travels to Claude. Google never hears you. Anthropic never hears you. Nobody hears you.
| Step | What Happens | Where |
|---|---|---|
| PTT PRESSED | Physical button activates microphone. Nothing else triggers audio capture. | On device only |
| AUDIO CAPTURED | Raw audio recorded in device memory. Does not leave the device. | On device only |
| WHISPER STT | Audio sent to your Mac Mini over local WiFi. Converted to text. Audio discarded immediately. | Your local network only |
| TEXT TO CLAUDE | Text transcript sent to Anthropic API over HTTPS. Audio is never sent. | Text only — HTTPS encrypted |
| PIPER TTS | Claude response converted to speech by Piper on your Mac Mini. No third party. | Your local network only |
| AUDIO PLAYED | Voice plays on device speaker. Audio file discarded after playback. | On device only |
| VOICE NOTES | Audio encrypted by TDLib before leaving. Carrier never hears content. | E2E encrypted to recipient |
| REMINDERS | AlarmManager fires locally. Piper speaks reminder. Never leaves device. | On device only — always |
No audio. No voice. No vocal fingerprint. No behavioural profiling. Nothing that identifies your voice ever leaves hardware you own.
— Text of your query — words only, no audio
— Timestamp of the request
— Your API key — not your name or identity
— Your IP address — mitigated by VPN if desired
Anthropic never receives your voice, your vocal characteristics, your location, your name, or your phone number.
— Your voice or vocal fingerprint — ever
— Ambient audio — microphone is hardware-gated
— Conversation history — each query is stateless
— Advertising inference data — no profiling
— Identity — no name, email, or number required
Nothing that identifies your voice, face, or body ever leaves hardware you own.
"The central problem with smartphone use today is you have no idea what the hell it's doing at any given time."
— Edward Snowden, Joe Rogan Experience
Snowden spent his career inside the world's most powerful surveillance apparatus. He described exactly this problem in detail. Telegotchi is the answer he said didn't exist.
No ambient recording. No voice fingerprinting. No advertising inference. No corporate server has ever heard your voice.
That is not a feature. It is the architecture.