Stage audio tool-routing adaptation and use typed argument heads

For audio-conditioned tool routing with a small ASR language model, add task-specific route and typed argument heads to the fused audio-text hidden state and adapt the backbone in stages. Start with the backbone frozen; train pooling, route, presence, enum, and span/candidate heads. Then add decoder LoRA and train the audio projector. Unfreeze upper audio-encoder layers only when call-disjoint validation shows audio-domain errors that the lighter stages do not fix. Preserve recognition with an auxiliary ASR loss and replay data. This applies to fixed or versioned tool schemas; open-ended values such as prices, IDs, dates, and names should use transcript-grounded span/candidate heads or constrained decoding rather than unrestricted regression.

Limits: model-specific hidden-state access may require a wrapper, and route labels must match the exact deployed tool schema. A small routing dataset is adequate for a head-only feasibility test but does not justify broad encoder fine-tuning.

Applies when: Training a compact audio-text model to emit calibrated tool routes and structured arguments on limited domain data. Keywords: audio tool routing Qwen ASR fine tuning heads arguments calibration