TNT 🧨

🌐 appautomaton.com/tnt-asr — the project landing page.

Terminal voice-to-text. Tap Space, speak, tap Space — your words land in the transcript and on the clipboard.

Qwen3-ASR-1.7B runs in-process on the Apple GPU via mlx-speech as an 8-bit (int8) quantized checkpoint — ~2.5 GB resident: the model loads once, stays resident, and transcribes a short take in a fraction of a second. Fully local — no cloud, no runtime network calls. The microphone is captured natively through AVFoundation by a small Swift helper process, so a misbehaving audio stack can never trap the mic: TNT just kills the helper and macOS releases it.

Features

• In-process GPU inference — pure MLX, no PyTorch
• 8-bit quantized — int8 weights (~2.5 GB), about half the memory of BF16 with a faster decode
• Resident model — loads once in the background at startup; every take is warm
• Native mic capture — AVFoundation via an isolated Swift helper process; the mic can always be reclaimed
• English, Chinese, and mixed speech — language auto-detected, or forced via env var
• Live braille oscilloscope — real audio levels while you record
• Clipboard-first — new transcriptions auto-copy; click any past entry to copy it again
• Responsive TUI — side-rail layout on wide terminals, stacked on narrow ones

Setup

git clone https://github.com/appautomaton/tnt-asr.git
cd tnt-asr
uv sync
./bootstrap-mlx-asr.sh        # downloads + links the int8 checkpoint (~2.5 GB, cached by Hugging Face)
uv run tnt

Or install from PyPI (automaton-tnt):

uv tool install automaton-tnt
TNT_MLX_MODEL=/path/to/qwen3-asr-1.7b-int8-mlx tnt

(Instead of exporting TNT_MLX_MODEL, you can symlink the checkpoint at ~/.local/share/tnt/qwen3-asr-mlx.)

Model checkpoint

TNT expects a converted Qwen3-ASR-1.7B MLX checkpoint. A ready-to-use int8 build (~2.5 GB) is published at appautomaton/qwen3-asr-1.7b-int8-mlx. The bootstrap script takes three forms:

./bootstrap-mlx-asr.sh                       # download the int8 build from Hugging Face, then link it
./bootstrap-mlx-asr.sh <hf-repo-id>          # download a specific Hugging Face repo
./bootstrap-mlx-asr.sh /path/to/checkpoint   # link a checkpoint you already have (no download)

Downloads use huggingface_hub (already installed via mlx-speech) and land in the shared Hugging Face cache (~/.cache/huggingface); the script symlinks bin/qwen3-asr-mlx to the cached snapshot. It is idempotent — if the model is already cached, or you pass a local path, nothing is re-downloaded, so you never keep two copies of the 2.5 GB weights. BF16 and mxfp8 builds work too — mlx-speech reads the quantization from the checkpoint's config.json, so switching is just a relink. Alternatively, convert the upstream Qwen/Qwen3-ASR-1.7B weights yourself with mlx-speech's scripts/convert/qwen3_asr.py.

Configuration

Keybindings

Project structure

src/tnt/
├── app.py             # Textual TUI, state machine, keybindings
├── audio.py           # Recorder protocol, backend selection, PortAudio (non-macOS)
├── avf_audio.py       # Native AVFoundation capture via helper process (macOS)
├── mic_helper.swift   # AVFoundation helper source, compiled on demand
├── async_threads.py   # Daemon-thread helpers for blocking work
├── transcriber.py     # In-process MLX Qwen3-ASR transcription
└── widgets/
    ├── transcript.py  # Scrollable transcript log
    └── status.py      # Braille oscilloscope + state rail
bin/
└── qwen3-asr-mlx      # Symlink to converted MLX checkpoint (gitignored)

Related projects

• mlx-speech — our MLX-native speech runtime that powers TNT (PyPI)
• qwen3-asr-1.7b-int8-mlx — our int8 MLX checkpoint that TNT runs (converted from Qwen3-ASR-1.7B)

More from appautomaton

• 🌐 appautomaton.com — our site
• 🤗 huggingface.co/appautomaton — our models and checkpoints on Hugging Face
• 🐙 github.com/appautomaton — our open-source projects

License

MIT. See LICENSE.