Recently, I wrote about an issue with Hungarian, and now we have a problem with the pronunciation of the Norwegian language. When listening to the language on 11labs, everything sounds good, with clear pronunciation of all words. Recording: [Google Drive Link](https://drive.google.com/file/d/1bKNgZ_2CADIOdLhKwHnS0DJ2kB7WMPmj/view?usp=sharing) However, on a Vapi call with the same model (Eleven Flash v2.5), we hear something entirely different, especially at the end of the phrase. Recording: [Vapi Call Recording](https://storage.vapi.ai/f1c068ea-4677-4ec2-9447-d329452af9cb-1742387009899-ebac2a96-3144-4b47-a878-fc2102e9bf03-stereo.wav) - Vapi call ID: f1c068ea-4677-4ec2-9447-d329452af9cb - Model: Eleven Flash v2.5 (We also tried Turbo 2.5, no difference) - Voice ID: b3jcIbyC3BSnaRu8avEk - Bergen - Test phrase Hei! Takk for at du ringer Sorrissi, jeg heter Erik. Hvordan kan jeg hjelpe deg i dag? We need Norwegian without a Danish accent, as heard in the 11labs recording. Why does it sound correct on 11labs but not on the platform? Please help, this is a serious issue. I previously encountered a similar problem related to the Hungarian language. https://discord.com/channels/1211482211119796234/1346436410495729664

@Karina Krinitskaya Sorry for your experience. I have shared your Google call recording, VAPI call recording, and other details with the 11labs team and FBI. For this call, you are using a multilingual model, not Flash model, and your assistant current 11labs model is turbo.

Response from Inlama Labs team sharing here as it is for fair consideration and transparency. https://cdn.discordapp.com/attachments/1351895867648507954/1352274345929605231/Screenshot_2025-03-20_at_7.04.50_PM.png?ex=67dd6b20&is=67dc19a0&hm=875d964814b026faa9b5567e8cb847fca29a9bb324557941dccca1a9885b591c&

Eleven Labs*

Hi

Hey Karina, For the shared call ID, the model which was used is a multilingual model. Below I'm sharing the logs which you are sending to 11labs stream. You can take a look. It's not Flash model, and they have also confirmed the same. 🔵 12:42:31:876 ElevenLabs (WebSocket #0) Connecting... URL: wss://api.elevenlabs.io/v1/text-to-speech/b3jcIbyC3BSnaRu8avEk/stream-input?output_format=pcm_16000&optimize_streaming_latency=3&model_id=eleven_multilingual_v2&enable_logging=false&enable_ssml_parsing=true 🔵 12:42:32:096 Getting Speech For

11labs:b3jcIbyC3BSnaRu8avEk:0.5:0.75:undefined:undefined:undefined:eleven_multilingual_v2:16000:c3273a0022fe23fb110bd32a214f97b391b12bfbeeed1a6676b36ea774e0199d

, "88" Text: ""Hei! Takk for at du ringer Sorrissi, jeg heter Erik. Hvordan kan jeg hjelpe deg i dag?""

@Shubham Bajaj Hey Let me jump into the conversation — Karina and I double-checked this case. A call with call_id

e4eefb33-aeca-4cee-ad2f-6299aeb12a30

was made using the

7de2b370-01c3-4bd7-9af5-cfc7e92c338b

assistant directly on the VAPI. The assistant’s settings explicitly specify the

Eleven Flash V2.5

model, and I can also retrieve this model via the request (we’re also taking into account that information on your platform is often displayed incorrectly):

GET https://api.vapi.ai/assistant/7de2b370-01c3-4bd7-9af5-cfc7e92c338b

200 OK
Response body: {
    "id": "7de2b370-01c3-4bd7-9af5-cfc7e92c338b",
    ...
    "name": "vapi_caller",
    "voice": {
        "model": "eleven_flash_v2_5",
        "voiceId": "b3jcIbyC3BSnaRu8avEk",
        "provider": "11labs",
        "stability": 0.5,
        "similarityBoost": 0.75,
        "enableSsmlParsing": true,
        "inputReformattingEnabled": false,
        "inputPreprocessingEnabled": true
    },
   ...
    "firstMessage": "Hei! Takk for at du ringer Sorrissi, jeg heter Erik. Hvordan kan jeg hjelpe deg i dag?",
    ...
}

I understand that the logs on your side show something different, but we’re also unable to get the expected result through your platform. So clearly, there’s some issue on your end — please look into this matter as soon as possible.

@august @Karina Krinitskaya I just reviewed the shared call ID. Your current modal config is FlashModal, but it's falling back to Multilingual, which is why it's being used during the call. I'm check with the team why it's changing to Multilingual?

Hi! Sorry, we are still waiting for your response. :)🐣

Hey Karina, Sorry for taking time to get back to you. The issue has been fixed. There was a bug because of legacy code it was falling back to the multilingual model, and when I was going through the logs, it showed me that you guys are using the multilingual model despite the current model being the Flash model. Earlier, when the Flash model was released, the only English language transcription was possible, and we added a check so that users never run into call failures. Now, we have removed the check, so you can start using the Flash model.

@Shubham Bajaj Double-checked — it’s working now, thank you!

@Shubham Bajaj Thank you so much! \[T]/

Marking this ticket as Solved ✅