turn-classifier-mmbert-v1

Multilingual Turn Classifier mmBERT v1

Private five-language, four-head textual turn-taking classifier for SIP voice agents. The selected checkpoint is jhu-clsp/mmBERT-base seed 17, exported as ONNX and dynamically quantized to signed INT8.

Runtime contract

Input template:

[LANG] <it|en|fr|es|de>
[STATE] <LISTENING|ASSISTANT_PLAYING|ASSISTANT_THINKING>
[HISTORY]
<agent|user>: <text>
[USER] <current user transcription>

Maximum input length is 256 tokenizer tokens. The four outputs, in fixed order, are:

1. endpoint: COMPLETE | INCOMPLETE
2. floor_claim: CLAIM | BACKCHANNEL
3. addressee: AGENT | SIDE_TALK
4. call_intent: CLOSING | NORMAL

manifest.json contains the runtime temperatures, thresholds, label order, and checksums. Consumers must verify it before loading model.int8.onnx.

Release 1.0.0

• INT8 model: 308,090,421 bytes
• FP32/INT8 argmax agreement: 99.25%
• warm CPU p95: 46.48 ms on the release benchmark host
• blocking golden cases: 24/24
• held-out INT8 metric regression: passed
• challenge SIDE_TALK precision: 98.32%

The full immutable audit bundle is stored at the repository root. The runtime model is model.int8.onnx; model.fp32.onnx and the JSON reports are retained for parity, provenance, and audit.

Loading

Use the versioned source repository's turn_classifier.infrastructure.onnx_runtime.OnnxClassifier with manifest.json. The model is intended for local CPU inference and does not send conversation text to an external service.

Limitations

• Supported languages are limited to Italian, English, French, Spanish, and German.
• call_intent/CLOSING is diagnostic and should not trigger automatic hangup without an application policy.
• SIDE_TALK prioritizes precision over recall to avoid suppressing replies to speech addressed to the agent.
• Latency depends strongly on sequence length and target hardware.