speak — local text-to-speech CLI

Speak text out loud from the command line using Supertonic 3 — a fast, ~99M-parameter TTS model that runs entirely on your CPU. No cloud, no API keys.

speak "hello world"
echo "piped text works too" | speak
speak -v noah "a different voice"
speak "مرحبا بالعالم"          # language auto-detected from the text
speak -o clip.wav "save to a file instead"

Install

uv tool install speak-cli

Or without waiting for a PyPI release, straight from GitHub:

uv tool install git+https://github.com/MohamedAliRashad/tts-cli

(From a clone, uv tool install . works too.) All dependencies are prebuilt Python wheels — no compilers, no apt packages. On the first speak, the Supertonic 3 models (~400 MB) are downloaded once to ~/.cache/supertonic3/; everything after that works fully offline. say is installed as an alias of speak (handy, but if another tool on your machine already provides say, just use speak).

How it stays fast

The first call starts a background daemon that keeps the models loaded in memory, so subsequent calls speak in well under a second. The daemon exits after 15 minutes idle (tune with SPEAK_IDLE_TIMEOUT seconds) and is respawned transparently. If the daemon can't run for any reason, speak silently falls back to in-process synthesis — it always works.

speak --stop        # stop the daemon manually
speak --no-daemon   # bypass the daemon for one call

Voices

speak --list-voices

--voice accepts any of these names, case-insensitive. Set a persistent default with export SPEAK_VOICE=noah.

Options

Language auto-detection is script-based: Arabic, Japanese, Korean, Russian, Greek, and Hindi are detected from their alphabets; all Latin-script text is assumed English (use --lang fr, --lang de, … to override). Supported languages: en ko ja ar bg cs da de el es et fi fr hi hr hu id it lt lv nl pl pt ro ru sk sl sv tr uk vi.

Expression tags can be embedded in the text: speak "well <laugh> that was funny" (also <breath>, <sigh>, …).

Playback

Audio is played through the first working system player among paplay, pw-play, aplay, afplay (macOS), ffplay, play (sox) — every mainstream desktop has at least one. If none works, speak tells you and suggests --out.

Licenses

CLI code: MIT. Supertonic 3 model weights: OpenRAIL-M (via the supertonic SDK's auto-download).