The AI can hear the caller, but the caller hears silence. A connected call with one-way audio points to a media-path problem that should be investigated before changing the response prompt.
For engineers diagnosing voice reception, separate telephone signalling from audio transport. A successful connection event establishes that a call was accepted, not that both directions of media work.
What is one-way audio in a voice AI call?
One-way audio means one participant receives usable sound while the other does not. The failure can occur between the carrier, SIP service, real-time room, agent input or agent output.
First determine the missing direction. Compare what the caller hears, what arrives at the agent and whether the agent generates output. This avoids changing unrelated components because every symptom has been labelled silence.
Trace the media path with one controlled call
Use an authorised test call and retain the relevant identifiers. Limit access to packet captures and recordings because they can contain sensitive call information.
| Observation | Next boundary to inspect |
|---|---|
| Caller speech never reaches SIP service | Provider route and incoming media |
| Incoming media exists, agent gets no input | Subscription and room-to-agent path |
| Agent generates no response | Session, turn and generation events |
| Agent publishes audio, caller hears nothing | Outgoing media and destination path |
| One direction changes after transfer | Transfer-specific route and media negotiation |
These are investigation branches, not guaranteed diagnoses. Preserve the evidence that supports moving from one boundary to the next.
Inspect negotiated addresses and actual packets
The LiveKit SIP troubleshooting guide describes how media addressing and network translation can cause one-way audio. Compare the negotiated media destination with the path actually receiving packets.
A private or otherwise unreachable advertised address is a useful clue in an applicable deployment. Managed and self-hosted arrangements have different responsibilities, so confirm which component owns that address before changing configuration.
Do not broadly open network access as a diagnostic shortcut. Apply the documented requirements for the actual service and validate the narrow change through another controlled call.
Confirm the repair across the affected route
Retest the original direction and the reverse direction, including the transfer state if that triggered the failure. A working browser call does not verify the same telephone media path.
My voice platform contribution provides application context. The forwarding verification guide helps keep the public route and fallback in the acceptance record.
Record what was observed and what remains untested. A single fixed destination does not establish that every provider or carrier route uses the same media behaviour.
Does a connected call prove the audio path works?
No. Inspect media delivery in both directions separately from the signalling result.
Should we change the AI model first?
Check whether input reaches the agent and output leaves it before attributing silence to generation.
What should we provide for diagnosis?
Share anonymised identifiers, the missing direction and relevant events through the voice AI project contact.