I called the Telnyx number I'd connected to my Vapi assistant. The agent answered and I could understand its voice. When I spoke, the audio reaching the agent was so distorted that even a person couldn't understand it. There was no useful transcription, and the agent stayed silent.
The call had two audio directions. The greeting travelled from the assistant to my phone. My reply had to travel back through the telephone connection to the speech-recognition service. Hearing the greeting established that one direction worked; it didn't establish that my voice arrived in a usable form.
The two audio directions behaved differently
assistant → phoneClear audioThe call connected and the caller could understand the agent.
caller → recognitionDistorted audioThe caller's speech was unintelligible at the agent side.
I had imported the number through Vapi's Telnyx integration panel. I started testing the audio formats offered in the connection settings. These are codecs: the rules used to encode and decode the sound.
The three tests
Telnyx's assistant suggested adding Opus. I enabled it and called again. The distortion remained.
Next I left G711U as the only selected audio codec. VP8 and H.264 were still selected as video codecs. That configuration made the connection worse: the call disconnected, Vapi showed “unable to decode audio data” and I couldn't hear the assistant anymore.
I then restored G722, G711A and G711U, the original three audio choices. The call connected again and the assistant's voice returned. My incoming audio was still distorted.
| Change I tested | Connection | My voice reaching the agent |
|---|---|---|
| Add Opus to the original setup | Connected | Still distorted |
| G711U audio only, VP8/H.264 still selected | Disconnected | Decode error |
| Restore G722, G711A and G711U | Connected again | Still distorted |
That last test recovered the call, not the original problem. It would be misleading to call it a successful audio fix just because the greeting returned.
A plausible explanation still needs a recording
Vapi's Telnyx guide describes a failure that could fit: some calls outside North America can arrive using G.711 A-law while the native integration expects G.711 µ-law. Interpreting one encoding as the other can produce garbled caller audio and incomplete transcripts. The guide also distinguishes this telephone path from browser calls, which use WebRTC.
That is a hypothesis for this investigation. I haven't established which encoding was actually negotiated on my calls.
The suggested next diagnostic was to compare the caller recording at Telnyx with the caller recording at Vapi and inspect the negotiated media format. I don't have that completed comparison. The last observed state remains a connected call, clear assistant speech and an unusable caller track.
Project sources & checks
User reports describe clear assistant audio, unintelligible caller audio, the unsuccessful Opus test, a G711U-only audio configuration that disconnected and restored connectivity after the original codecs were re-enabled.
The native Telnyx guide describes a possible A-law/µ-law mismatch on some calls outside North America. That explanation has not been confirmed for these calls.
No successful repair, completed comparison of the two providers’ caller recordings or packet-capture result is established.