I have a very simple prompt (trimmed down to test this problem): "# Role You are a helpful and knowledgeable virtual assistant for a company. # Conversation Flow - Ask for their first and last name. - Repeat their name slowly and ask them to confirm. - Ask for their phone number. Let them provide their full phone number. Do not interrupt until they have provided their full phone number, which is 10 digits. - Repeat their phone number slowly, digit by digit, and ask them to confirm." In about 30%+ of the calls the AI ignores the names of the caller and repeats the question. It's especially bad with less generic names such as Austin. Example calls: fe8f771b-4a9d-4dbb-aa1d-07bf5cb799e4 9951f913-aaf2-414b-8c44-58b0b83755a5 In the message log I see that the transcriber had output the name correctly. I've tried setting the "endpointing" for deepgram to anything between 100 and 300 Tried "smartEndpointing", played with various values for the "Start speaking plan" settings, tweaked the prompt following any of the online samples. Even tried the Vapi Blocks beta. Nothing helps, the failure rate for that specific question stays around 30-50%, which is terrible.

@olivaw I'm sorry to hear that you're experiencing issues with the voice recognition system. It seems like you've already tried several adjustments to improve the accuracy. Since you mentioned that Deepgram does not support your language and you've been using the Talkscriber Whisper model, it's important to ensure that the model is correctly configured to handle the nuances of the names it's struggling with. Here are a few steps you can take to potentially improve the performance: 1. **Custom Pronunciation Dictionary**: If the names are consistently being misunderstood, you might consider implementing a custom pronunciation dictionary if the service allows it, to help the model better recognize and transcribe unique or less common names. 2. **Adjust Prompt Wording**: Sometimes, rephrasing the prompt can lead to better recognition. For example, you could ask for the caller's name in a different way that might be clearer for the model to understand. 3. **Post-Processing Logic**: Implement logic that checks the transcribed text for common errors and attempts to correct them based on context or a predefined list of names. 4. **Feedback Loop**: Use the instances where the AI fails to understand the names as data points to train the model further, if the service provides a mechanism for continuous learning. 5. **Technical Support**: Since you've tried many solutions without success, it might be time to reach out to the technical support team for the voice recognition service you're using. They might have insights or solutions specific to their system that could help resolve the issue. 6. **Testing with Different Models**: If possible, test with different models or services to see if another might provide better accuracy for your use case. If you'd like, I can assist you in reaching out to the technical support for further help. Would you like me to initiate that process for you?

1. This is not feasible 2. I've tried anything, from adding extra context to combining with the phone number in one question. I'm open to suggestions, but nothin I found here or online seems to help 3. These are real-time phone calls plus they seem to be transcribed fine? 4. How? This is based on GPT 4o Cluster