Feature: Enable transcription language selector in Room settings (currently commented out)

[spatialy]

Summary

The RoomLanguageSelector component is fully built and functional — supporting 45 languages with flag emojis and native labels — but the UI row is commented out in RoomTranscriptionSettings.svelte, making it inaccessible to users. All rooms default to language: 'en' at creation time with no way to change it from the UI.

This means:

• OpenAI STT always receives language: "en", producing garbled output and low confidence scores (~30%) for non-English speakers
• Deepgram STT is unaffected by user choice (hardcoded to language: "multi" auto-detect, updateLanguage() is a no-op)

Provider-specific language behavior

Even if the UI selector were enabled today, it would only affect OpenAI STT. Deepgram's implementation intentionally ignores the room language:

Behavior	OpenAI STT	Deepgram STT
Language default	"en" (stored on instance)	N/A — hardcoded "multi" in getOptions()
updateLanguage() behavior	Stores new language, sends live transcription_session.update to all active WebSocket connections (gated by OPENAI_PROVIDE_LANGUAGE)	No-op — method body is entirely commented out
Language sent to API	this.language (dynamic, from room metadata) or undefined if OPENAI_PROVIDE_LANGUAGE=false	Always "multi" (hardcoded), regardless of room setting
Auto-detection	Only when language is omitted (OPENAI_PROVIDE_LANGUAGE=false)	Always — Nova-3 "multi" mode handles multilingual automatically
Mid-meeting language change	Works — sends transcription_session.update in real-time	Would require connection restart (code exists but is commented out)

The Deepgram updateLanguage() originally had a stop/restart implementation to reconnect with the new language, but it was commented out — likely because Nova-3's "multi" mode already handles multilingual meetings well without explicit hints.

The infrastructure is already complete

Layer	Status	Details
Types	✅ Done	RoomLanguage union type with 45 languages, languagesDisplayData with flag emojis
Room schema	✅ Done	Room.language: RoomLanguage field persisted in DB
UI component	✅ Done	RoomLanguageSelector.svelte — dropdown with search, two display modes
Settings popup	❌ Commented out	RoomTranscriptionSettings.svelte lines ~34–39
Room creation default	❌ Hardcoded 'en'	AddRoomPopup.svelte — ignores browser locale
Metadata flow	✅ Done	startTranscription() → connectMeeting(room.language) → aibot → love → LiveKit metadata
Love service	✅ Done	/transcription and /language endpoints update room metadata
Love-agent	✅ Done	RoomMetadataChanged → stt.updateLanguage()
OpenAI STT	✅ Done	transcription_session.update with new language, works dynamically mid-meeting
Deepgram STT	❌ No-op	updateLanguage() body is commented out, getOptions() hardcodes 'multi'

Where the code is commented out

plugins/love-resources/src/components/RoomTranscriptionSettings.svelte:

<!-- <div class="antiGrid-row">
    <div class="antiGrid-row__header">
      <Label label={ui.string.Language} />
    </div>
    <RoomLanguageSelector {room} />
  </div> -->

plugins/love-resources/src/components/AddRoomPopup.svelte:

language: 'en',  // hardcoded, no browser locale fallback

services/ai-bot/love-agent/src/deepgram/stt.ts:

// private language: string = 'en'

updateLanguage (language: string): void {
  // const shouldRestart = (this.language ?? 'en') !== language
  // this.language = language
  // if (shouldRestart) {
  //   this.stop()
  //   this.start()
  // }
}

Impact on self-hosted deployments

Self-hosted users deploying with OpenAI STT in non-English environments get:

• Low confidence transcriptions (30–50% probability)
• Garbled text for non-English speech
• No way to fix it without modifying source code or switching to Deepgram

Proposed fix

1. Uncomment the language selector row in RoomTranscriptionSettings.svelte and add the missing imports (ui plugin reference and RoomLanguageSelector)
2. Default to browser locale in AddRoomPopup.svelte instead of hardcoded 'en' (fallback to 'en' if the browser locale is not in the supported list)
3. Restore updateLanguage() in deepgram/stt.ts — uncomment the stop/restart implementation and wire this.language through getOptions() (default 'multi' preserves current behaviour; setting a language overrides auto-detect for single-language meetings)

Workarounds (until fixed)

• Switch to Deepgram (STT_PROVIDER=deepgram) — uses language: "multi" auto-detect, works for all languages
• Set OPENAI_PROVIDE_LANGUAGE=false on the love-agent — disables the English hint, lets OpenAI auto-detect (less accurate than an explicit language hint)
• Direct DB update — UPDATE love SET data = jsonb_set(data, '{language}', '"es"') WHERE _class = 'love:class:Room' (requires CockroachDB access)

Environment

• Version: v0.7.353 (hardcoreeng/* images)
• STT Provider: OpenAI (Realtime API, gpt-4o-transcribe)
• Deployment: Self-hosted via Docker Compose

[huly-github-staging]

Connected to Huly®: UBERF-15624