bargein-classifier

bargein-classifier

Open-source barge-in detection for voice agents. Classifies whether a user is interrupting (barge-in) or just backchanneling ("uh-huh", "yeah") during an agent's turn.

Designed as a self-hosted alternative to proprietary adaptive interruption handling.

Overview

How It Works

Place downstream of VAD in a voice pipeline. When the user speaks while the agent is talking:

1. VAD detects speech overlap
2. This model scores the user's audio
3. High probability = barge-in, agent should yield
4. Low probability = backchannel or noise, agent keeps speaking

The model expects the user's audio in isolation (after echo cancellation), not a mixed signal.

Performance

Detection accuracy

Hard negative rejection

Cross-corpus generalization (AMI-only variant)

Quickstart

ONNX Runtime

import numpy as np
import onnxruntime as ort

session = ort.InferenceSession("bargein.onnx", providers=["CPUExecutionProvider"])
meta = np.load("bargein.onnx.meta.npz", allow_pickle=True)
threshold = float(meta["threshold"][0])

# features: (1, 1, 64, 200) log-mel spectrogram from 2s of 16kHz audio
logits = session.run(None, {"input": features})[0]
prob = 1.0 / (1.0 + np.exp(-logits[0]))
is_bargein = prob >= threshold

HTTP Server

BARGEIN_MODEL_PATH=bargein.onnx uvicorn server.app:app --host 0.0.0.0 --port 8080

curl -X POST http://localhost:8080/bargein --data-binary @audio.raw
# {"is_bargein": true, "probability": 0.87, "threshold": 0.339, "prediction_duration_ms": 4.8}

Deployment

Training Data

~55K labeled events from two meeting corpora:

• AMI (138 meetings, ~29K events) — dialog-act annotations as weak supervision
• ICSI (75 meetings, ~26K events) — MRDA dialog-act tags
• Hard negatives (~5K) — laugh, cough, throat clear, breath sounds

Audio source: individual per-speaker headset channels, analogous to a user's microphone with echo cancellation.

Labels are weak supervision from dialog-act ontologies, not human-audited barge-in judgments.

Limitations

• English only — trained on English meeting corpora
• Domain gap — trained on meeting audio, not voice agent audio. Deploy in shadow mode first to validate on your traffic.
• Single-speaker input — expects isolated user audio (with echo cancellation). Performance degrades on mixed/summed channels.
• Weak labels — there is a label noise ceiling from the dialog-act proxy. Human-audited fine-tuning would improve quality.

Training Configuration

Author

Borislav Novikov ([email protected])

Citation

@software{novikov2026bargein,
  title={bargein-classifier: Open-source barge-in detection for voice agents},
  author={Novikov, Borislav},
  year={2026},
  url={https://huggingface.co/bnovikov/bargein-classifier}
}

Training data

@article{carletta2005ami,
  title={The AMI meeting corpus: A pre-announcement},
  author={Carletta, Jean},
  journal={Machine Learning for Multimodal Interaction},
  year={2005}
}

@inproceedings{janin2003icsi,
  title={The ICSI meeting corpus},
  author={Janin, Adam and Baron, Don and Edwards, Jane and Ellis, Dan and Gelbart, David and Morgan, Nelson and Peskin, Barbara and Pfau, Thilo and Shriberg, Elizabeth and Stolcke, Andreas and Wooters, Chuck},
  booktitle={ICASSP},
  year={2003}
}

License

Model weights: CC BY 4.0 (same as training data).