[feat] Server-wide default language for live transcription

[bakiburakogun]

Is your feature request related to a problem? Please describe.

The live transcription language is stored per conversation only. LiveTranscriptionService::enable() reads $room->getLiveTranscriptionLanguageId() and omits langId entirely when it is empty, so the live_transcription app falls back to its hardcoded default (TranscribeRequest.langId: str = "en").

On an instance where the spoken language is not English this means:

• every conversation starts in English until somebody opens the conversation settings and changes it,
• users who do not know about that setting see English transcription of non-English speech, which produces nonsense,
• the obvious-looking workaround — picking their own language in the personal "translation target language" setting — makes it worse, because the speech is still transcribed with the wrong acoustic model and is then machine-translated on top, adding a task-processing round trip per segment.

We run a Nextcloud for a public-sector organisation in Türkiye with several thousand users. Turkish is the language of essentially every call. There is no way to express that once for the whole instance.

Describe the solution you'd like

A server-level default that is used when a conversation has no explicit language set, for example:

occ config:app:set spreed transcription_language_default --value tr

LiveTranscriptionService::enable() would fall back to that value before falling back to the app's own default, and the conversation setting would continue to override it. The conversation settings dropdown could show it as "Default language (Türkçe)" instead of the current hardcoded "Default language (English)".

An admin UI field in the Talk settings would be nice but the occ/app-config key alone would already solve the problem for us — it is scriptable, which is what matters for a large deployment.

Describe alternatives you've considered

• Setting the language on every conversation by hand. Not workable at our size, and it does not help conversations created later.
• Updating oc_talk_rooms.transcription_language directly in the database. It would fix existing rooms but bypasses the service layer and still leaves new rooms defaulting to English.
• Changing the default in the live_transcription app instead. That app already carries a # todo: declarative settings for language override note, but the value logically belongs to Talk, since Talk is what decides whether to send langId at all.

Additional context

Related: the same "make the default configurable instead of hardcoded" gap exists for the Photos and Talk attachment folder names, which we also hit on this deployment.

Nextcloud version: 34.0.3
Talk version: 24.0.4
live_transcription version: 2.1.3

Happy to open a PR for the app-config fallback if the approach sounds right to you.

[Antreesy]

Need a backender or maintainer opinion (cc @kyteinsky @nickvergessen), but I feel that if live_transcription app is a provider, and have limited number of available languages to transcript from, default language should be defined/configured there (same place where it's hardcoded TranscribeRequest.langId: str = "en"))
Talk should also receives a list of available languages (I assume, also from live_transcription itself), so it feels like it belongs there

[bakiburakogun]

That makes sense to me, and it lines up with a TODO already sitting in the app: ex_app/lib/main.py carries # todo: declarative settings for language override and model download.

For it to work end to end I think it needs two pieces:

live_transcription — make the default behind TranscribeRequest.langId configurable (an env var alongside the existing LT_* ones, validated against VOSK_SUPPORTED_LANGUAGE_MAP) and report it in /capabilities under live_transcription.default_language, next to the supported_languages already exposed there. Nothing else has to change for transcription itself: LiveTranscriptionService::enable() omits langId entirely when the conversation has no language set, so the app's own default is what takes effect today.

Talk — src/components/ConversationSettings/LiveTranscriptionSettings.vue hardcodes Default language (English). With the capability above it could render the actual default instead, otherwise the setting would read "English" while the app transcribes in something else, which is the confusing half of what we ran into.

Worth noting the failure mode this produces on a non-English deployment, since it is quiet: the transcription starts in English, the user sees nonsense, and the obvious-looking fix — picking their own language in the personal "language used to show the transcriptions" setting — makes it worse, because the speech is still decoded with the wrong acoustic model and then machine-translated on top, adding a task-processing round trip per segment.

Happy to open a PR for the app side, and a follow-up for the label in Talk, if that split sounds right to you.