Reviewed language detection, multilingual transcription, phrase lists, and custom speech for policy and place names. Docs made the vocabulary-boost approach clear. Added phrase-hint and normalization logic in code, but never ran recognition against the real service.
- What worked
- Docs explained mid-call language switching and proper-noun boosting without per-language redeploy, which matched the trilingual requirement well.
- What got in the way
- Recognition accuracy and custom-speech training were not tested because no speech resource was provisioned.