I use Retell to run about 30 calls a day. About 1 in 30 face disconnection issues. These are almost always logged as user hangups, but when you listen to the call, it’s clear that the agent and the person stopped hearing each other, and eventually the disconnection led the person to off, triggering the hang up. These are almost always calls via the API (never the phone). Sometimes the disconnection happens in the middle of the call. Other times the audio seems not to work from the beginning.
For the life of me, I can’t figure out what’s causing this. Any ideas?
From our investigation, the Retell server pipeline (ASR, LLM, TTS) is healthy on these calls; the audio loss originates on the your side (browser/WebRTC), after which the call ends with a user hangup.
Hey Shaw — Our team has investigated thoroughly on our side. Even if the server pipeline is healthy, the evidence suggests audio is dropping at the LiveKit media transport layer, which we cannot address since this is managed entirely by the Retell SDK.
Our integration follows ReTell’s documentation exactly. We don’t touch the LiveKit room, ICE/TURN, remote track attachment, or media reconnection, and our backend returns 200 OK on every call creation.
The transcripts we shared point in the same direction:
- One call shows the user saying “Hello?” three times. The agent responds “I can hear you — can you hear me?” The user says “Hello?” again. ASR is transcribing speech fine, but agent audio isn’t reaching the browser. One-way audio loss at the transport layer.
- Multiple calls show users cut off mid-word and the agent prompts “still there?” repeatedly with no response. Same pattern.
- We saw one guy try 5 web calls in a single day, each with the same result. They never completed their interview.
We use the same agents, LLM, and prompts for phone and web calls. This doesn’t happen over the phone, ruling out Claude streaming gaps or LLM-side issues. It points to web call transport.
For the call IDs we provided, would you be able to share WebRTC session diagnostics? That would help us better nail down whether this is user-environment or transport-layer.
To be direct, this is causing real problems for my clients. At ~$10k in monthly spend, I need a path to resolution here.