Hello, Based on migration guide from docs to realtime models, I should remove the transcriber configuration, but when I do so, the transcriber parameter in assistant code is missing but it sets default to deepgram EN, also in end-of-call report. Should I set the transcriber in realtime models to match language and get transcript in end-of-call report? Will realtime handle multilungual transcripts in future (with openAI transcribers)? Thank you. https://docs.vapi.ai/openai-realtime https://cdn.discordapp.com/attachments/1485187379643027509/1485187379848417361/Screenshot_2026-03-22_at_8.58.21.png?ex=69c0f408&is=69bfa288&hm=987d7153251a91ab1893832bc506561e9a9319a4e4e83bd6b6357c01c5b0e398&

You should remove the transcriber configuration when using OpenAI realtime models, as these models natively handle speech-to-speech and do not require a separate transcriber setting. To migrate, update your assistant configuration like this:

json
{
  "model": {
    "provider": "openai",
    "model": "gpt-realtime-2025-08-28"
  }
}

No need to set a transcriber for language matching; transcripts are generated by the realtime model itself. For multilingual transcript support, the documentation does not mention future plans—current limitations apply. See the [OpenAI Realtime migration guide](https://docs.vapi.ai/openai-realtime) for details. Source: - [OpenAI Realtime Documentation](https://docs.vapi.ai/openai-realtime)

When I turn transcriber off, the assistant can handle the call on backend, but the transcriber is I think default set to depgram, see callID 019d14e9-a4d7-7000-9d63-b2ef6fb5e381, where the transcriber is almosst missing, but the assistant handled the call in czech language. Is there a way to let the realtime handled the transcriber, so I get than work with it in end-of-call report etc?

Yeah, you’re on the right track, when you remove the transcriber, VAPI just falls back to Deepgram (usually EN), which is why the report doesn’t match the actual language. If you care about clean transcripts in the end-of-call report (especially for Czech), it’s safer to explicitly set the transcriber to the right language for now. Realtime can run the conversation fine, but transcripts and reports still rely on that config. Multilingual support on OpenAI realtime is getting better, but it’s not fully there yet for consistent reporting. Out of curiosity, are you using the transcripts more for logs, analytics, or triggering workflows after the call? @honza.r

you should still configure a transcriber even with OpenAI Realtime models. Vapi uses it separately for endpointing, interruption detection, and generating end-of-call transcripts. Without it, it defaults to Deepgram Nova-3 EN, which is why you're seeing English in your reports. To get transcripts in the correct language, explicitly set it:

Hi @Ramsey thank you for response. I was thinking the same, but just wanted to confirm, because of the documentation. I am using the transcript in end-of-call report for clients and for a workflow where next automations are going based on transcripts. Will set transcriber as I need, but I like the multilingual idea without squad multiagent architecture in the future, as @Shaunak said. Thank you both guys.

No worries! If you face any further issues, feel free to create a new ticket.