Hey all! Can someone help me understand what might be going on here? The agent's voice sounds great, but the caller audio coming into the agent is extremely bad. In the web test, my mic sounds fine, but on real phone calls it's quiet, muffled, and barely understandable. The transcription struggles a lot because of it. I've tried both a Vapi-provided number and my own imported number, and the issue is still there. It honestly feels like 1/10 audio quality, not just normal phone compression.

Hi there, Thank you for your message. Our team is currently out of the office. We operate Monday through Friday, from 9:00 AM to 8:00 PM Pacific Standard Time (PST). We’ll get back to you as soon as possible during our normal business hours. If your message is urgent, please mark it accordingly or include “URGENT” in the subject line, and we’ll do our best to respond promptly. Warm regards, Vapi Customer Support Team

This usually points to the telephony audio chain, codec, or noise suppression settings rather than the agent itself. I’ve seen cases where PSTN routing, Twilio/Vapi configs, or carrier-level compression completely destroy inbound caller audio and tank STT accuracy. I can help you debug it directly and pinpoint where the degradation is happening. Are you using Twilio behind Vapi, and which STT provider/model are you currently running? @cheezyyye

Hi, can you please share call id so we can investigate more

Thanks. My provider is Vonage, but I also tried using a Vapi-provided number directly and the quality issue is identical, so it does not seem isolated to Vonage. The main thing is that the web test sounds good, but real phone calls sound terrible. In the actual Vapi recordings, the caller side is much quieter than the assistant, so the STT is struggling because the raw inbound caller audio is so quiet and muffled. I also tried enabling diagnostics like wav;l16 artifacts, logs, and pcapEnabled, but that did not resolve it. Call IDs: Vonage number: 019e3b66-e503-7666-aeb4-de3379dcaac7 Vapi-provided number, different assistant, same issue: 019e3b6d-50ee-7001-89dd-c29471f45b24

Hi, Thank you for sharing the details. As checked from our end, both calls sounded perfectly fine, and the user transcription is also visible in the logs. As a recommendation, we suggest using Deepgram Flux transcription in your configuration and setting the language to

en

for better transcription performance and call quality consistency.

I do not agree that this is solved. I have tried everything, and the problem is very clear. I am using the same microphone in both tests. The web call sounds fine, but the real phone call sounds extremely bad. I understand that phone calls have some compression, and I expect that. However, this is not normal phone compression. This is a severe failure in how the audio is being transferred. I am attaching the waveforms here as a direct comparison. You can clearly see the difference. The phone call audio is heavily degraded, much quieter, and muffled compared to the web call, even though the same sentence was spoken using the same microphone. This is also confirmed by your own agent struggling to transcribe the caller audio properly. If the transcription cannot reliably understand clear speech from a good microphone, then the call quality is not acceptable. In its current state it simply does not work. https://cdn.discordapp.com/attachments/1505760600448565389/1507448986770866308/comparison.png?ex=6a11f0c5&is=6a109f45&hm=a29886f78eae8dc7801b23aa764a5771d4ed7fa2e7abc514a3bac31744cab1bf&

Hi, a few things I would like to acknowledge and confirm. 1. As you already know, voice calls need some compression, and it's a lot in comparison to web calls, like phone calls work on 16k or 8k sample rate. 2. The two call IDs you shared look alright to me. One thing I noticed is that the volume of the user is lower than the assistant's voice, but that can be accounted for by the generationConfig volume = 1.6, so the assistant's voice is set to be louder. I am not claiming they are perfect, but I am claiming they don't have any major issues. 3. You mentioned the transcriber didn't work, can you please point to the call ID, and the timestamp where the transcriber failed to transcribe the user's voice? I checked roughly for both call but wasn't able to catch it. 4. Is it possible to try Twilio once? AFAIK the quality is better in Twilio because of more native websockets implementation.