✨ What it does

Two 150M models, ~12 ms per decision on a laptop GPU (~30–60 ms on CPU). English, text in, text out: plug them in after any speech-to-text.

🚀 Install

pip install fennec-voice

The GPU is used when available, otherwise the CPU. For a smaller CPU-only install, get PyTorch's CPU build first:

pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install fennec-voice

🧩 Use

from fennec_voice import FennecInterrupt, FennecHandoff

interrupt = FennecInterrupt()
r = interrupt.predict(said="Your flight from Mumbai to Delhi is",   # what the caller had heard
                      unsaid="on the seventeenth of October.",      # the rest of the agent's sentence
                      caller="mm hmm")                               # what the caller said
r.label, r.keep_talking        # ('BACKCHANNEL', True)
r.p_stop                       # 0.002: probability of "stop and listen"
r.probs                        # {'BACKCHANNEL': 0.998, 'ACK_AFTER_QUESTION': ..., ...} all 10 labels

handoff = FennecHandoff()
h = handoff.predict("I want to fly from Mumbai to um")
h.finished, h.p_unfinished     # (False, 0.99): wait for more
handoff.predict("Delhi, one way for two people.")   # HandoffResult(finished=True, p_unfinished=0.011): reply

• said / unsaid: split the agent's sentence at the word playing when the caller started speaking (TTS word timestamps + playback position).
• predict_batch([...]) runs many inputs in one pass.
• Results have .to_dict().

Labels (interrupt.labels): BACKCHANNEL (a nod: keep talking) and, for stop and listen: ACK_AFTER_QUESTION, ANSWER, QUESTION, CORRECTION, STOP_REQUEST, HOLD_REQUEST, CANT_HEAR, PICKUP_GREETING, TAKES_TURN.

⚙️ Options

FennecInterrupt(
    model="latest",      # or "fennec-interrupt-150m-v1", a Hugging Face repo id, or a local folder
    device="auto",       # "cuda" or "cpu"
    fp16=None,           # default: on for GPU
    threads=None,        # CPU threads
    revision=None,       # pin exact model files (Hugging Face commit or tag)
)
FennecHandoff(..., threshold=0.5)   # p_unfinished at or above it means "not finished"

import fennec_voice
fennec_voice.list_models()   # all known models

The package and the models have separate versions: new models arrive as new names, and "latest" only moves to a better default. Pin a name for stable behaviour.

▶️ Demo

fennec-voice-demo: a full local voice agent built around these two models, one ./run.sh.

💡 Good to know

• English only.
• "okay" can mean "I'm listening" or "yes": right before the agent's yes/no question it counts as "yes".
• v1, a hobby project: expect rough edges.

📄 License

Apache-2.0. The input formatter comes from ozonetg/bargein-classifier-ettin-150m (MIT), see NOTICE.