Small synthetic fine-tunes don't fix a turn model's pause errors

A pretrained full-duplex turn-taking model that cuts in during mid-number dictation pauses was not fixed by threshold recalibration, retraining only its heads, retraining its GRU controller, or adding dialogue-state inputs, all trained on a small synthetic two-channel corpus. Every gain in catching real interruptions also brought false stops on acknowledgments like "mm-hm", and no variant learned to hold during pauses. Transcript-level guards on top of the stock model, plus a semantic decider, worked better. Frozen self-supervised audio encoders (JEPA-style, WavLM) with small heads trained on mostly non-English turn data also collapsed on English domain speech. Limits: the corpus was tiny and TTS-generated. Real consented call audio could change the result.

Applies when: deciding whether to fine-tune an open turn-taking or endpointing model for domain-specific pauses. Keywords: turn taking fine-tune Tyger Nemotron endpoint barge-in backchannel dictation pause synthetic data JEPA WavLM