Agent continually repeats and ignores interactions...
# support
d
Works great in the web dashboard but all over the place via Twilio, any ideas? I am using Groq/Llama3 and for STT I am using Nova2. 9dbe19ee-9e48-4951-afdc-934f5173664e
v
Hi @dynjo, I'm an AI assistant for Vapi. I'll analyze available sources and get back to you shortly.
----------------------------- **While I generate an answer, here are some sources that might be helpful**: * [List Calls]() * [Get Call]() * [Call Ended Reason]() * [Create Phone Call]() -----------------------------
### Answer The behavior you're experiencing with your agent when using Twilio, as opposed to its performance on the web dashboard, might be attributed to differences in how audio is processed in web environments versus over phone lines. One known issue is that the performance of Speech-to-Text (STT) models, like Nova2, can vary significantly between these environments due to differences in audio quality and transmission methods [(1)]() . For English, Deepgram offers a fine-tuned model specifically for phone calls, which might not be the case for all languages or models. Since you're using Nova2 from Deepgram, it's possible that the model's performance is optimized for web audio rather than the compressed audio formats used in phone systems. This could explain why the agent repeats or ignores interactions - it might be struggling with accurately transcribing the audio from Twilio phone calls. A possible solution mentioned in the forums is to try different STT providers or models that might have better performance over phone calls. For instance, switching to a different STT engine like Whisper might yield different results. However, it's crucial to experiment with different configurations to find the one that best suits your needs. Additionally, ensure your Twilio integration settings and the way you're handling audio streams align with Vapi's best practices. Misconfigurations could also lead to issues like the ones you're experiencing. If you continue to face challenges, considering the specific use case and the detailed behavior you described, directly reaching out to Vapi's support with the call IDs and detailed explanation could also provide more tailored assistance.
s
@dynjo Lower models tend to hallucinate a lot. To fix this, I suggest moving to GPT-4. If there are latency and cost issues, consider fine-tuning the GPT-3.5 model for use.
d
@Sahil but the exact same config works perfectly through the web UI, so this seems to be a problem with the interaction between Twilio and Vapi?
v
Got it. Tagging @User for additional help.
s
Please try once with GPT-4 and let me know if the issue persists.
d
It does seem better with GPT-4