When spelling a phone number like “514 729 0234,” the system incorrectly interprets it as “54 72 90 0.” It seems that the STT tries to associate the number with the French (France) phone number format, leading to errors. Example: User: “Mon numéro de téléphone est le cinquante-quatre sept-deux-neuf, zéro.” Bot: “Donc j’ai votre numéro de téléphone. Cinquante-quatre soixante-douze quatre-vingt-dix zéro.” This issue is currently putting all our assistants on hold.

@Simon J. I understand the issue you're facing with the speech-to-text interpretation of phone numbers. It seems like there might be a need to adjust the language model or the parsing logic to correctly interpret the sequence of numbers as intended for the phone number format. To address this, we can explore the following options: 1. **Language Model Adjustment**: We can look into whether the speech-to-text service offers customization options for language models, allowing us to prioritize certain interpretations over others. 2. **Post-Processing Logic**: Implement a post-processing step in your application that can recognize and correct the common misinterpretations based on the expected phone number format. 3. **Explicit Context**: Provide explicit context to the STT service if it supports it, indicating that the expected input is a phone number, which might help it to parse the sequence correctly. 4. **User Correction**: Implement a feature that allows users to correct the phone number if it is misinterpreted, ensuring that the correct number is captured. If you need assistance with implementing any of these solutions or if you have further questions, please let me know, and I'll be happy to help.

I ran several tests using the Deepgram API, and here are the results for different scenarios: French: • Nova-2: Same Issue • Enhanced: Same Issue • Base: Same Issue • Whisper Cloud: It works French (Canada): • Nova-2: Same Issue • Base: Same Issue • Whisper Cloud: It works Can we enable the Whisper Cloud model for Deepgram STT in VAPI?

can you share call id for this.

thanks for the reply, here the call id : 2bad1f5e-64ab-41a3-a06c-1dd680f8b844

i will get back to you on this what can be done.

thanks Shubham!

@Simon J. try using deepgram keywords, and phonecall model.

@Shubham Bajaj how do we define the Deepgram keywords setting in VAPI? Phonecall model is not supported for the French language https://cdn.discordapp.com/attachments/1268930482867798108/1270063793644507226/image.png?ex=66b25695&is=66b10515&hm=ee2082ae5739a3a9e7178a52e761d0e043771417e67fd4e7920ff24e89cc25c8&

- avoid phone-call model - ask LLM to generate numbers as words in french language and add these number words as keywords in deepgram

hi @Shubham Bajaj , i have same problem with italian. Im using Nova as STT, azure as TTS and gpt 4o as LLM, when i try to tell during a call a identy card number like AC12345BB its really hard to let the model understand all the characters, do you have any tips?

@Shubham Bajaj Is there an option for custom STT yet?

@Alejo not yet, but soon will be made available.