# Voice, your microphone and your privacy

When GPTree listens (including in a voice session), where your speech is turned into text, what Live voice sends to OpenAI, what GPTree keeps (your words, never recordings), how spoken replies are made, how long your words are kept, and how to see, delete or stop keeping them.

The short version: GPTree listens only when you start it and shows it the whole time, and it never stores recordings of your voice. With Standard voice, your speech is turned into text before anything reaches GPTree. With Live voice, your voice streams to OpenAI so it can answer as you talk. The words you said are kept for 30 days unless you switch that off, and you can see or delete them in Voice history.

## When the microphone is on

Voice starts only when you start it: the mic in the top bar or the waveform button in the message box (a voice session), a tap or hold on the mic in the message box, or the push-to-talk keys.

- **In a voice session**, you can keep talking after each reply, even while an answer is still being written in the conversation. The session stays on until you end it, mute it, switch to another tab or app, or go quiet for about a minute, when it pauses itself and the mic closes. A paused session ends by itself after a few minutes. The voice bar is on screen the whole time the mic is on.
- **After a single command**, it keeps listening for a few seconds so you can follow up; the bar shows "Listening for a follow-up" with a countdown for as long as that lasts.

It stops at once when you press Esc, click End or Stop, say "stop" or "end session", or switch to another tab or app. There is no always-on listening.

## Standard voice: where your speech becomes text

- **In a browser**, the browser's own speech recognition turns speech into text. Depending on the browser, that happens on your computer or through the browser maker's speech service. The Voice panel says which: "transcribed on this device", or that your browser's speech service transcribes.
- **In the GPTree app for Mac and the iPhone app**, recognition runs on the device. If your language has no on-device model, voice says it is not available rather than sending your audio to a server.
- **In the Android app**, the phone's speech recognition service is used, on the device when the phone supports it.

When a browser uses its maker's speech service, that service receives audio for as long as the mic is on, which in a voice session includes the pauses between turns.

With Standard voice, what reaches GPTree is the text of what you said, plus the names and ids of what is on your screen so "that answer" can be matched to the right message. Audio is never sent to GPTree.

## Live voice: what goes to OpenAI

Live voice is a realtime voice from OpenAI that hears you and answers as you talk, while your own AI model writes the detailed answers in your conversation. It comes with voice on every plan that includes voice. Voice sessions use Live voice whenever it is available; when it is not, or your browser or device cannot run it, the session uses Standard voice instead. If you would rather your voice did not stream to OpenAI, turn off **Speak replies** in the Voice panel: voice sessions then use Standard voice, with replies shown as captions.

- **Your audio** streams from your device to OpenAI for as long as the Live session is on, including the pauses between turns. GPTree's servers receive the same stream alongside it, so a session can be billed by the second and ended when your credits or the session time run out. GPTree never stores or logs that audio.
- **OpenAI** is asked not to keep a recording of the session for later use. OpenAI's API data terms still apply, which can include keeping data for a limited time to check for abuse.
- **The screen:** the live voice is told the names of what is on your screen, such as the conversation title and the model answering, never the text of your messages.
- **Answers:** when your conversation's answer is ready, its spoken version is sent to OpenAI as text for the live voice to say.

## What is kept, and for how long

- **Your words:** what you said, kept for 30 days so you can review it, then deleted. With Live voice this includes what the voice said back.
- **What voice did:** a short record of each action, such as "Opened Q4 pricing strategy", so the session list and receipts work.

During a session, voice also holds the last few things you said and its replies for up to three minutes after you last spoke, so a correction or a follow-up makes sense. That short memory is never saved, it ends with the session, and it works the same with Keep what I said off.

To see a session's words, open **Voice history** in the Voice panel; you can delete any session there, which forgets its words. A session also shows as a Voice session card in the conversation it happened in, and only you see your own sessions, even in a shared conversation.

To stop keeping your words, turn off **Keep what I said** in the Voice panel. From then on only the record of what voice did is kept, and Voice history shows when each turn happened without the words. Your voice history is included when you export your data, and it is deleted with your account.

## Spoken replies

With Standard voice and Speak replies on, replies and summaries are read aloud by a natural voice. To make that audio, the text of the reply is sent to OpenAI's voice service; your own audio is never sent, and the audio made from the reply is not stored. The audio uses credits. To keep replies on your device instead, turn off **Natural voice** in the Voice panel and your device's built-in voice reads them. Everything spoken is also written out on screen. Turn Speak replies off, or tap CC on a phone, for captions only.

With Live voice, the live voice itself does the speaking, so Natural voice applies to Standard voice only. Live voice always speaks its replies: with Speak replies off, voice sessions use Standard voice with captions.

A spoken summary of a long answer is written from the answer's text by the fast AI model voice uses for commands. With Standard voice the summary itself is not stored; with Live voice, what the voice says is part of the words kept under Keep what I said. A short, plain answer is read as it is, with no summary.

## If you have blocked the microphone

If you said no to the microphone earlier, GPTree cannot ask again. Allow it in your browser's site settings for GPTree, or for the GPTree app in your phone's or your Mac's privacy settings, then try again. Voice tells you when this is the problem.
