Joint semantic route heads can outperform scalar generation, but exact identifiers remain a separate bottleneck

General lesson: for voice tool calling with a bounded schema, train a tool-qualified semantic-route classifier jointly with an auxiliary tool head, while keeping price and reference extraction in a deterministic post-route normalizer. This applies when tool/action classification is strong but a learned scalar decoder remains unreliable. A flat route catalog avoids ambiguous duplicate enum strings across tools, and an auxiliary tool loss improves the shared representation. Positive-pair consistency is useful; a hard-negative margin may add no value when sampled negatives already satisfy the margin, so its actual activation rate should be reported rather than inferred from its configured weight. Limits: high synthetic validation accuracy does not establish production generalization, and deterministic extraction remains bounded by ASR quality.

Applies when: designing a speech-to-tool decoder with bounded enums plus prices or identifiers. Keywords: speech tool calling semantic routes scalar extraction ASR LoRA margin