I’m validating a tool - not-model approach for teams running local Whisper pipelines:

• Manage domain glossaries (terms/aliases/weights) → export glosspacks.

• Optional: train a tiny bias adapter on your audio+list; drop-in wrapper (logit fusion).

• A/B eval with WER + rare-word metrics.

For those shipping Whisper in production:

How do you handle rare words/terminology today?

Do you bias at decode time today? How?

What pack/adaptor formats would fit your stack?

Which failure modes worry you (over-biasing, latency, multi-lingual)?

Would love your feedback. Thanks!