Hi everyone, I’m running into a strange issue in WAPI with an Assistant during a Test Call. What’s happening: -In the Test Call, I can hear the assistant voice normally (TTS audio sounds fine). -But the transcript is 100% gibberish / nonsense: -Assistant → Transcript: The assistant’s spoken output is displayed as random text. -User → Transcript: When I speak, my speech is also transcribed as nonsense. Since the model seems to rely on the transcript, it then “understands” only nonsense and the conversation can’t continue. My configuration Provider: OpenAI Model: GPT 4o Cluster First Message Mode: Assistant speaks first Voice Configuration Provider: OpenAI Voice: Onyx TTS Model: Tts 1 UI warning shown: “You are using de language for your transcriber. Make sure you choose a voice compatible with de.” Transcriber Provider: OpenAI Model: GPT-4o Transcribe Language: German Background Denoising: Enabled auto), etc. Any known cases where TTS is audible but ASR transcripts become gibberish? Happy to share screenshots of my voice/transcriber configuration if that helps. Thanks! 🙏 https://cdn.discordapp.com/attachments/1459168209679683700/1459168210300702874/image.png?ex=697b584b&is=697a06cb&hm=19e965902a3e85c1b65598d227a4b50457d0c1928047dc62ec67f07a5afa96ff&

Even in English I've seen occurences where the bot clearly understands the works base on its reaction but in the transcription they are incorrect. I think they are using a separate LLM for the transcript that is dumber but lets see what the support says and is it possible to use the same LLM as the bot for the transcript too

Hi, thanks for the details. This usually happens due to a language/voice mismatch. You've set the transcriber language to German, but if the audio/voice isn't actually German, ASR can output gibberish. Could you please share the assistant ID so we can confirm the exact config. Thanks!

sure the assistants id is ab713788-f31a-4054-b051-b89a5984a0e3

Thanks for sharing! The issue was caused by a language mismatch between the voice and the transcriber. Using an English oriented voice (OpenAI Onyx) with a German transcription settings can cause the transcript to appear as gibberish. For example, this config should work correctly:

"voice": {
  "voiceId": "de-DE-KillianNeural",
  "provider": "azure"
},
"transcriber": {
  "model": "nova-2",
  "language": "de",
  "provider": "deepgram"
}

You can choose other providers if you prefer, just make sure the voice and transcriber both support the same language. Thanks, Shaunak

Hi Shaunak — understood. My main question isn’t which voice is compatible; it’s how I’m supposed to tell inside Vapi. In the Vapi UI, when I select Provider = OpenAI, I only see a short list of voices (Alloy / Echo / Fable / Onyx / Nova / Shimmer). My transcriber is configured for German (language = de). However, I don’t see any UI indicator that tells me: 1. whether a given voice supports German, 2. whether it’s “English-oriented” vs multilingual, 3. or whether it’s compatible with my current transcription language/model. So the mismatch you described makes sense — but I’m missing the discoverability: Where in Vapi can I verify voice language support / compatibility before running calls? Concretely: Is there a label/metadata in the dashboard (or API) that indicates each voice’s supported languages/locales? If not in the UI, is there a Vapi docs page or an endpoint that returns voice capabilities (supported languages) per provider? What’s the intended workflow for users here — trial-and-error, or is there a canonical compatibility reference? If you can point me to where this is surfaced (UI/docs/API), I can self-serve these checks going forward. Thanks!

You’re right. At the moment, Vapi does not clearly show voice language support or compatibility in the dashboard. There is also no API endpoint that lists supported languages or locales for each voice. Right now, the expected workflow is: • Check the original provider’s documentation (OpenAI, Azure, ElevenLabs) for language support. • Test the voice with your configured transcriber to confirm it works as expected. For German specifically, Azure voices are the safest option because they are explicitly language‑specific (for example de-DE-*). OpenAI voices are generally multilingual, but this is not currently indicated in the Vapi UI. We agree this is a discoverability gap, and your feedback has been shared internally. For now, the provider documentation is the best reference rather than a Vapi-side compatibility list. For reference: https://docs.vapi.ai/customization/multilingual

Thanks alot!