Voice typing in the prompt box

Voice typing writes what you say into the prompt box. You press the microphone, speak, press it again, and the words appear where your cursor is. Nothing is sent. You read the text, fix it if needed, and press Send yourself.

It is for long prompts, for describing a bug while you look at it, and for moments when your hands are busy. It is not a replacement for checking what came out.

Partly built. Voice typing works, but it makes mistakes, most of all with the names of tools, files and commands. We have not measured it on English speech, so look at the text before you send it.

Dictate a prompt

1. Click in the prompt box where the text should go.
2. Click the microphone beside the paper clip. The box says The microphone is opening; do not speak yet…
3. When it says Speak, talk. A level bar and a count of seconds show that the sound arrives.
4. Click the microphone again, now titled Done speaking. The box says Writing what you said…
5. The text appears in the box, with a line saying which engine wrote it and asking you to look at it before you send.

If the microphone gives no sound, the box asks whether it is muted or whether another one is chosen. A recording with no speech in it writes nothing and says so. If a take fails, the recording is kept and a button, Write it again, tries it once more without making you repeat yourself.

The microphone is open only while you record, and the studio lets go of it when you stop. If Windows blocks it, the message points you to the Windows privacy setting for microphone access by desktop apps.

First use: the local engine

By default the speech is turned into text on your own computer, with no internet and no cost, and your voice never leaves it. The first time you press the microphone, a panel says it needs two files: the whisper program (about 8 MB) and its model (about 574 MB). Press Download both, or, if you already have the model file, write its path and press Use this file. The studio only accepts the model file it expects. Later closes the panel. The first-run setup lets you choose this engine, but the download itself waits for your first press of the microphone; see First-run setup.

The local program runs only while it writes a take. When you are not dictating, no model sits in memory.

Choose the engine: your computer or OpenAI

Open Settings, then Speech to text under Writing and sending.

• whisper on this computer (free) is the default.
• OpenAI with your key sends the recording to OpenAI and is paid by the minute on your own key. Add the key first under Paid keys, then choose the key and a model; each model in the list shows its price per minute.

The cost is checked against any spending limit you set for keys before the recording is sent. If the key is refused, the credit is gone, or OpenAI cannot be reached, the box says which. When the credit runs out, the studio tells you, offers Charge it and Look now, and until you charge it your voice is written by the local engine, if that is set up. It looks again by itself every half hour. A ChatGPT subscription cannot be used here; speech needs an OpenAI key with credit.

Words of each project

Speech engines stumble on the names a project uses. So each project has its own list, Words of this project, one word to a line. The studio fills it from the project's own files: the folder name, package.json, Cargo.toml, requirements.txt and the names of git branches, plus a list of common programming words. The list is given to the engine as a hint, so these words are more likely to come out as you spell them.

Open the list from the line under a take, or from the setup panel, edit it in the box and press Save. A word you remove is not gathered back. The list is a plain text file, and the box shows its path.

Voice typing writes in the language your sessions reply in, which is English until you change Language of replies in Settings. To talk to the studio itself instead of typing into a session, see The voice assistant. The rest of the box is in The prompt box.

Small tips

Questions