feat(deepgram): support flux-general-multi and language_hint in STTv2

[robertvy]

Feature Request

Please add flux-general-multi and repeatable language_hint support to the Deepgram STTv2 plugin.

Why this belongs in the plugin

livekit-plugins-deepgram already implements the Deepgram v2 streaming API directly over WebSocket. It is not blocked on Deepgram SDK support for this feature.

Relevant plugin files:

• livekit/plugins/deepgram/stt_v2.py
• livekit/plugins/deepgram/_utils.py
• livekit/plugins/deepgram/models.py

Today the plugin already supports Flux v2 for English-only use cases via flux-general-en, but it does not expose the Flux multilingual path.

What is missing today

1. flux-general-multi is not included in the plugin's v2 model set.
2. language_hint is not exposed on STTv2 constructor or update_options.
3. language_hint is not forwarded as repeated query parameters on the v2 WebSocket URL.
4. Auto-detected language information is not surfaced cleanly for Flux multilingual responses.

Requested support

1. Allow model="flux-general-multi" in STTv2.
2. Add language_hint: list[str] | None support on:
   • constructor
   • update_options(...)
   • stream reconnection / URL generation
3. Encode hints as repeated query params on /v2/listen, e.g.:
   • ...?model=flux-general-multi&language_hint=de&language_hint=en
4. For flux-general-multi, use returned language metadata from the API when emitting transcript events instead of defaulting to a fixed plugin language.
5. Optionally expose mid-stream reconfiguration for Flux multilingual if that fits the plugin API.

Relevant Deepgram docs

• Flux Multilingual & Language Prompting
  https://developers.deepgram.com/docs/flux/language-prompting

• Getting Started with Flux
  https://developers.deepgram.com/docs/flux/quickstart

The key points from those docs:

• flux-general-multi uses the same production v2 endpoint as Flux.
• language_hint is only supported on flux-general-multi.
• Hints are repeatable and optional.
• Without hints, the model auto-detects the spoken language.
• An EU endpoint is available for the multilingual model as well.

Why this matters

Flux multilingual is valuable for multilingual voice agents because it brings the same turn-aware / interruption-aware behavior as flux-general-en to supported non-English languages.

That means users can get Flux-style turn detection without staying on Nova-3 solely for multilingual traffic.

Implementation notes

This looks relatively contained:

• _utils._to_deepgram_url(...) already uses urlencode(..., doseq=True), so repeatable language_hint query params fit naturally.
• STTv2 already connects directly to /v2/listen over WebSocket.
• The main plugin work appears to be:
  • model typing / validation
  • option plumbing
  • event parsing for detected language metadata

Concrete example

stt = deepgram.STTv2(
    model="flux-general-multi",
    language_hint=["de", "en"],
    eot_threshold=0.7,
)

or with auto-detect:

stt = deepgram.STTv2(
    model="flux-general-multi",
)

Thanks.

[tinalenguyen]

hi @robertvy, i created a PR to support flux-general-multi with language_hint and its language detection feature (accessible via SpeechData.source_languages). please feel free to try it out and let me know your thoughts 😸