Voice

Akili Code can speak. Answers, documents, and the narration in a Motion video — read out loud by a voice that lives on your computer, not on a server.

It costs nothing. There is no per-character price and no metering, because nothing leaves the machine: the synthesiser runs locally, and it keeps working on a plane.

Getting a voice

Nothing is bundled with the installer, so the first voice is a download.

/voice add alba

That fetches a 52 MB engine once, then the voice itself — usually about 63 MB. Both are checked against a known checksum before anything is installed. When it finishes, the voice introduces itself, so you hear it immediately.

The commands

/read is the one that earns its keep: a long answer, listened to while you look at the code it is describing. Code blocks and tables are skipped rather than spelled out — nobody wants to hear a closing brace pronounced.

The voices

Why the licence column is there

Speech models are trained on datasets with different terms, and they sit side by side in the same collection. Most of the voices above are public domain or CC0 — narrate a client's video with them and nobody minds. One, neema, is the only Kiswahili voice that exists, and its dataset does not spell its terms out; Akili Code flags it when you install it, and you should read the dataset's page before selling anything made with it.

We would rather show you a licence you did not ask about than let you find out later.

Narration in Motion

Give a scene a narration line and ask for it to be spoken:

/motion narrate the scenes

Every scene with a line gets a clip, and each scene's duration snaps to the length of its own sentence, so the visuals last as long as the words. Run caption_motion afterwards to realign the captions. Scenes where you attached your own recording are left alone — a voice-over you recorded yourself is never overwritten.

Bring your own engine

If you already run Coqui TTS — including a voice cloned from a sample of your own — Akili Code can speak through it:

/voice coqui tts_models/en/ljspeech/vits
/voice coqui tts_models/multilingual/multi-dataset/xtts_v2 my-voice.wav

We do not download Coqui, and we never will: it needs Python, PyTorch and a system phonemiser, and that does not belong inside a 12 MB installer. If it is on your PATH, this registers it as a voice like any other and /say, /read and Motion narration all use it.

A warning, shown when you register one: XTTS‑v2 is under the Coqui Public Model License, which permits non‑commercial use only — and Coqui closed in January 2024, so there is no longer anyone to sell a commercial licence. Fine for your own work; not for a client's video.

If nothing plays

Audio goes through whatever the machine already has — paplay, aplay or ffplay on Linux, afplay on macOS, the built-in player on Windows. If your audio goes somewhere unusual, name the player yourself:

AKILI_VOICE_PLAYER="ffplay -nodisp -autoexit"

The clip is always written to a file first, and the path is printed when playback fails, so you can open it yourself.