The conversation
One long session with the microphone open, so you can cut in mid-reply.
What this page covers
OutcomeUse speech, typing and interruption as one conversation rather than three modes.
Before you beginComplete the first-conversation lesson and keep a project active for voice.
Follow these steps
A conversation is one continuous session rather than a series of requests. From Start conversation until you stop it, Ginu holds the microphone open — including while it is speaking, which is the only reason interrupting works at all.
Speak, or type in the composer at the bottom. They are the same conversation and the transcript keeps both, so you can read back an exact detail while still talking. Type
/in the composer to open Commands.Cut in whenever you want. Ginu stops within one block of audio, hears what you said and answers that instead. A short “mm” is treated as backchannel and does not stop it; words do.
Example
Explain the next task. Wait—start with the risk, then the steps.
The correction changes the direction of the reply. Read the transcript to check that Ginu heard the wording you intended.
Choose how to take part
Speaker mute gives you text replies. Microphone mute stops Ginu hearing you. They are separate controls, so you can type a request and still hear its answer, or speak and read the reply.
Pause holds the conversation: nothing is heard, spoken or delivered while it is paused. Background work continues. Resume when you are ready to take part again.
Use the transcript to check an exact detail. When you know the action you want, type / in the composer to open Commands.
Follow the thought.
From your microphone to a reply you can interrupt. Scroll through the six stages below.
- 01 / 06
Your voice
The microphone stays open for the whole session, including while Ginu is speaking. There is no key to hold and no control to reach for — interrupting is just talking.
Start with a thought.
- 02 / 06
Speech detection
Every frame is scored for speech, so Ginu knows when you started. Ending your turn is a judgement rather than a stopwatch: a trailing silence, and whether the sentence sounds finished. A dangling clause waits for the rest.
Room to pause.
- 03 / 06
Local transcription
Your computer turns speech into text — no speech server, no upload. Long turns are transcribed in pieces as you talk, so the words appear while you are still speaking rather than all at once at the end.
About 18× real time in a warm, timed session on an M-series Mac; transcription throughput compared with audio duration.
- 04 / 06
The agent at work
Your turn is appended to one long-lived CLI session, so the project context it already had is never replaced. It can answer, or spawn a worker that runs alongside the conversation. An action needing permission stops and waits for you.
Your CLI account. Your boundaries.
- 05 / 06
A reply takes shape
The answer streams back a token at a time. Ginu cuts it at sentence boundaries and strips the parts written for the screen rather than the ear, so what is spoken is not a read-aloud of a markdown document.
A sentence at a time.
- 06 / 06
Back to your ears
Your computer turns the reply into speech, and writes it out in 43 ms blocks rather than as one finished file. That is what lets it stop mid-word when you cut in, instead of finishing the sentence it had already committed to.
43 ms is the playback block size, not a guarantee for every interruption. Speech synthesis runs locally; no speech-server round trip.
An illustrated route through Ginu. Scroll distance is not elapsed time.
Notes
- Speaker mute, microphone mute and Pause are three different things. Mute the speaker for replies as text, mute the microphone to stop it hearing you, and pause to hold everything while background work carries on.
- The orange microphone indicator stays on for the whole session, muted or not: muting drops frames in software rather than closing the stream.
Check
Try the speaker mute and microphone mute separately. You should be able to choose whether Ginu speaks and whether it hears you.