Feature Type
Would make my life easier
Feature Description
deepgram.STTv2 (the Flux v2 API wrapper) does not expose Deepgram's numerals option, even though Deepgram's Flux endpoint supports it for both flux-general-en (full support) and flux-general-multi (all supported languages except Hindi and Japanese) — see https://developers.deepgram.com/docs/numerals.
Without it, Flux transcribes spoken numbers as words ("five") instead of digits ("5"), which is a real problem for agents that read back balances, account numbers, dates, etc.
The v1 deepgram.STT class already supports numerals (and smart_format). The STTv2 wrapper's does not. This change looks similar to #5480 / #5486 which added support for language_hint.
I have a PR ready for this.
Workarounds / Alternatives
Switch to deepgram.STT (v1, e.g. model="nova-3") instead of STTv2/Flux, which loses Flux's semantic end-of-turn detection.
Additional Context
Relevant plugin files:
livekit/plugins/deepgram/stt_v2.py
Feature Type
Would make my life easier
Feature Description
deepgram.STTv2(the Flux v2 API wrapper) does not expose Deepgram'snumeralsoption, even though Deepgram's Flux endpoint supports it for bothflux-general-en(full support) andflux-general-multi(all supported languages except Hindi and Japanese) — see https://developers.deepgram.com/docs/numerals.Without it, Flux transcribes spoken numbers as words ("five") instead of digits ("5"), which is a real problem for agents that read back balances, account numbers, dates, etc.
The v1
deepgram.STTclass already supportsnumerals(andsmart_format). TheSTTv2wrapper's does not. This change looks similar to #5480 / #5486 which added support forlanguage_hint.I have a PR ready for this.
Workarounds / Alternatives
Switch to
deepgram.STT(v1, e.g.model="nova-3") instead ofSTTv2/Flux, which loses Flux's semantic end-of-turn detection.Additional Context
Relevant plugin files:
livekit/plugins/deepgram/stt_v2.py