TNT 🧨

Terminal voice-to-text for Apple Silicon. Tap Space, speak, tap Space. The text lands in the transcript and on the clipboard.

Models run locally on the Apple GPU through mlx-speech. No cloud, no PyTorch.

• Qwen3-ASR-1.7B (default): transcribes each take after you stop.
• Confucius4-R2T2 (optional): transcribes live while you speak.

Press m to switch models. Only the active model stays in memory.

Note

Using Termux on Android? Use the legacy legacy/android-termux-qwen0.6b branch.

Setup

Requires an Apple Silicon Mac, Python 3.13+, uv, and the Xcode command line tools (xcode-select --install) for the Swift mic helper.

git clone https://github.com/appautomaton/tnt-asr.git
cd tnt-asr
uv sync
./bootstrap-mlx-asr.sh   # link or download the Qwen3-ASR checkpoint
uv run tnt

Or from PyPI: uv tool install automaton-tnt.

Model checkpoints

TNT reads models from local paths only.

Model	Checkpoint	Location
Qwen3-ASR	qwen3-asr-1.7b-int8-mlx	TNT_MLX_MODEL, else bin/qwen3-asr-mlx, else ~/.local/share/tnt/qwen3-asr-mlx
Confucius4-R2T2	confucius4-r2t2-bf16-mlx	TNT_R2T2_MODEL, else bin/r2t2-mlx, else ~/.local/share/tnt/r2t2-mlx

./bootstrap-mlx-asr.sh /path/to/checkpoint links a Qwen3-ASR checkpoint you already have. With no argument it downloads the int8 build. For R2T2, symlink your checkpoint to one of the locations above.

Configuration

Environment variable	Default	Description
TNT_MLX_MODEL	see above	Qwen3-ASR checkpoint
TNT_R2T2_MODEL	see above	Confucius4-R2T2 checkpoint
TNT_MLX_LANGUAGE	auto	Chinese, English, or auto. Chinese keeps mixed Chinese/English speech from being translated to English
TNT_INPUT_DEVICE	system default	Microphone, by index or name

Keys

Key	Action
Space	Start / stop recording (or hold to record); cancels a running transcription
m	Switch model
c	Copy the last entry
click	Copy the clicked entry
x	Clear the transcript
q	Quit

Related

• mlx-speech: the MLX speech runtime behind TNT
• huggingface.co/appautomaton: our MLX checkpoints

License

MIT. See LICENSE.