AIUS can turn your voice into a prompt draft. Recording and transcription happen on your own computer — your voice is never sent to AIUS Light or to a model provider.

Requirements

Dictation uses whisper-cli, ffmpeg and ffprobe on your computer. On macOS install them with brew install whisper-cpp ffmpeg; on Linux install the equivalent packages (whisper.cpp and FFmpeg). The speech model — whisper.cpp’s multilingual small, 465 MB — downloads itself into the AIUS cache the first time you dictate, so the first transcription takes longer while it downloads; after that it is instant. The download is skipped on computers that cannot run the model well — under 8 GB of memory, under 4 CPU cores, under 2 GB of free disk, or a connection too slow to fetch the model quickly. Dictation then says why instead of installing, and simply tries again next time, so an upgraded computer (or a faster network) picks it up without any setup.

In the terminal

The prompt has one dictation control — the ◉ dictate chip at the right of the prompt bar. While the speech model downloads and warms up in the background (right after aius starts), the chip reads ◌ dictation… and is not clickable yet — give it a moment. Click it (or run /dictate) to start recording; the chip turns red and shows ● REC with a live timer against the two-minute limit. Click it again — or wait for the limit — to stop. AIUS transcribes the recording and drops the text into your prompt as a draft: review or edit it, then send it yourself. Stopping a recording never sends anything. Double-click the chip to pause and again to resume — the chip turns amber and reads ❚❚ while paused. Hold the chip for about half a second to open its side menu: pause/resume the recording, discard it, or pick a language. /dictate-pause and /dictate-cancel do the same from the keyboard.
The TUI while dictation records: a toast says Recording — run /dictate again to stop and transcribe, and the prompt shows ● REC 0:05 / 2:00

Recording in the terminal

The TUI with dictation paused: the prompt chip shows ❚❚ 0:07 / 2:00 in amber

Recording paused — the chip turns amber and the timer holds

The TUI with the dictated transcript inserted into the prompt as a draft, ready to edit before sending

The transcript lands as an editable draft

The first time, macOS may ask for microphone access for your terminal app — allow it. On Linux the recording uses PulseAudio, on Windows a named microphone device. If the terminal cannot reach the microphone, /dictate offers the browser recorder described below instead. Languages. Dictation detects the spoken language automatically across the model’s 99 supported languages. To pin one — useful for short or noisy clips — open the chip’s hold menu, run /dictate-language, or use the palette command Dictation language; the choice stays for later recordings.
The Dictation language picker in the TUI: a searchable list with Auto-detect selected above English, Chinese, German, Spanish and more

Picking a dictation language in the terminal

In AIUS Light

The conversation composer in a project’s AIUS panel has the same one-button control: a single microphone button whose icon and colour mirror the state — a mic at rest, a filled ● in pulsing red while recording, ❚❚ in amber while paused and ◌ while transcribing. Click it, allow microphone access when the browser asks, and speak — a level meter and a ● REC timer against the two-minute limit show while it records. Click again to stop and transcribe; double-click to pause and resume. The text replaces your current selection in the composer as a draft you can edit before sending. Hold the button (or right-click it) to open the same side menu — a language picker with small icon buttons for pause/resume ❚❚, discard ✕ and close ▾, each with a tooltip. Escape while recording discards it, and otherwise closes the menu.
The AIUS panel composer while dictation records: the microphone button shows a filled red dot, a level meter runs under the input, and the status line reads REC 0:03 / 2:00 · click to finish · double-click to pause

Recording in the AIUS Light composer

The AIUS Light composer with the single microphone button, and below it the open dictation menu: language picker with pause, discard and close icons

The composer in AIUS Light with its dictation menu open

Dictation in AIUS Light prefers aius running in a terminal on the same computer — that keeps your voice on your machine. When no local terminal answers, the composer falls back to AIUS’s own dictation service and says “Transcribing on the AIUS service…” so you always know where your voice goes. Dictation in AIUS Light is tested in Chrome and Edge; if your browser does not allow the loopback connection the composer needs to reach your computer, it uses the hosted service instead.

The browser recorder

When the terminal cannot reach your microphone itself — no permission, no audio stack — /dictate offers to open a small dictation page in your browser instead. It runs on your computer (served by aius, on http://127.0.0.1:8766), shows the same single-button control with pause, discard and language in its hold menu, and gives you an editable draft field. Press Insert draft to drop the text into the waiting AIUS prompt, or edit it first. Closing the terminal or the prompt invalidates the page’s insert link; reopen it with /dictate.
The local dictation page with a Dictate button, an editable draft area, an Insert draft button, and the note that audio is transcribed locally and never submitted automatically

The browser dictation page, opened from the terminal

Configuration

Defaults suit most computers; these environment variables adjust dictation when needed:

Troubleshooting

The dictate control does not appear in the TUI. Update aius to the current release and restart it — the chip also stays hidden while dictation is unavailable: missing whisper-cli/ffmpeg/ffprobe, or a computer below the model’s floor (see Requirements). /dictate always says exactly which case it is. “Local dictation is unavailable” in AIUS Light. No terminal on this computer answered. Start aius in a terminal on the same computer, then retry — or, where AIUS’s hosted dictation service is enabled, the composer uses it automatically and says “Transcribing on the AIUS service…” instead. “dictation tools missing” or a status that never turns ready. Install whisper-cli, ffmpeg and ffprobe — on macOS brew install whisper-cpp ffmpeg; on Linux install whisper.cpp and FFmpeg from your distribution or build whisper.cpp from source. Restart aius afterwards. Dictation says the computer needs more resources. The speech model download was deferred because the machine is under 8 GB of memory, has fewer than 4 CPU cores, has under 2 GB of free disk, or the connection is too slow for the 465 MB model. Nothing is installed; dictation retries automatically on the next attempt, so more memory, disk or a faster network makes it available without further setup. The first transcription takes a while. That is the one-time model download. Later transcriptions reuse the cached model and return in about a second for a short clip on a recent machine. macOS records nothing or recording fails immediately. Open System Settings → Privacy & Security → Microphone and enable access for your terminal app, then restart the terminal. The browser does not ask for microphone access, or said “denied”. Reset the permission for the site (the lock icon in the address bar → Site settings → Microphone → Allow), then click the microphone button again. On macOS also check the browser itself is allowed under System Settings → Privacy & Security → Microphone. Linux picks the wrong microphone or no audio. Set the recording source in PulseAudio/PipeWire as the default input, or point AIUS_DICTATION_DEVICE at the device name ffmpeg should use. Windows specifics. Dictation records and transcribes on Windows, but pause/resume is unsupported — the pause control has no effect there. If device detection fails, set AIUS_DICTATION_DEVICE to the microphone’s name as ffmpeg -list_devices true -f dshow -i dummy reports it. “No speech recognized” or a wrong-language transcript. Speak longer, clearer audio close to the microphone, and pin the spoken language (/dictate-language, or the hold menu on any surface) — auto-detection can miss on short or noisy clips. Two terminals are open and only one serves AIUS Light. The first aius owns the loopback port; the second keeps its own in-terminal dictation regardless. Dictating from the TUI works in either.

Privacy

  • Audio is transcribed on your computer by a local speech model; nothing about the recording leaves the machine.
  • Recordings are limited to two minutes and 20 MB, and the audio files are deleted once the transcript is made.
  • A dictated prompt is always a draft until you send it yourself.
  • For usage statistics AIUS counts only that a dictation happened — its surface (terminal, composer or browser page), duration, language, whether it succeeded, and the transcript’s length. The audio and the words are never sent anywhere.