T3 Voice command model
Say "switch to opus" or "show the terminal", and this model tells T3 Code which button to press. Anything that is not a command goes to the coding agent as normal.
It is Julia-1 (144M parameters), fine-tuned on the T3 Voice Commands dataset. It runs on CPU: about 40 ms per turn on 8 threads. Try it in the demo Space, or use it with t3-voice.
How it decides
The bar is set high on purpose: a wrong command is worse than a missed one, because a missed one still reaches the agent.
Result
On a test set that it never saw, written by a different writer (427 turns):
| Input | Commands caught | Wrong triggers |
|---|---|---|
| Typed text | 92.1% | 1.3% |
| Speech to text | 91.0% | 2.1% |
Most misses are bare answers like "nope" or "yeah". Without those, it catches 96.3% of commands.
Use
from huggingface_hub import snapshot_download
from t3_commands import CommandModel # in this repo
path = snapshot_download(
"yuvidhepe/t3-voice-julia1",
allow_patterns=["encoder_fp32.onnx", "heads.npz", "tokenizer/*", "t3_commands.py"],
)
model = CommandModel(path)
model.predict("switch to opus") # command change_model, argument claude-opus-4-1
model.predict("fix the failing test") # command none
Needs numpy, onnxruntime and tokenizers. No torch.
Limits
- English only, and turns of 12 words or less.
- It can still fire by mistake, about 1 to 2 times in 100 normal sentences. Give users an undo.
- Trained on synthetic sentences, not real speakers.
Labels, threshold, training setup and all numbers: DETAILS.md. License: Apache-2.0, like the base model.