Dear VAPI Support Team, We are experiencing a critical production issue with bidirectional audio transmission that is occurring frequently in our live calls. This is causing immediate business disruption as customers cannot hear our agent in real-time and the agent cannot hear customers, despite both appearing in post-call recordings. Organization ID: ac4993f3-edcf-484d-a00c-c654245a4a6d Issue Overview Symptoms: Client cannot hear agent speaking during live call Agent cannot hear client's voice during live call Post-call recordings and transcripts show complete conversation (both sides) This is happening frequently, not just isolated incidents Technical Analysis We compared a working call vs. problematic calls and found critical differences: Working Call (0199ab4d-ff1f-7668-88eb-3f9774085934): ElevenLabs: Normal voice processing Deepgram: Final transcripts contain client messages Result: Complete bidirectional audio Problematic Calls (Example: 0199ab28-821d-7337-93d2-684cca7bfa65): ElevenLabs: Agent message marked as "voice cached" Deepgram: NO final transcript of client speech Recording: Only agent voice appears, client audio missing Result: Complete audio transmission failure Additional Affected Call IDs: 0199aaf2-9f2c-7eeb-8cfe-8d1e89c1753d 0199a72b-af59-7ffe-bf98-deca3606a396 0199a6e4-b2a3-7772-9f7d-c3547d8e6038 0199ab1c-c988-722d-8f00-d5081d641fe0 0199ab1b-d323-7113-a168-0612caeab07a Key Findings ElevenLabs "voice cached" indicator appears in failing calls - suggests audio isn't streaming in real-time Deepgram shows no client audio transcription despite caller being connected - indicates audio pipeline failure Recordings contain full conversation - issue is specifically with real-time streaming, not recording System Configuration Integration: Go High Level + VAPI + n8n + Twilio Transcription: Deepgram Voice Synthesis: ElevenLabs Affected: Both inbound and outbound calls

Hi, I reviewed your report and it looks like the issue is with real-time audio streaming between ElevenLabs and Deepgram, not the recordings. I can help trace the media flow between VAPI and Twilio to see where the audio is dropping. The “voice cached” message suggests the agent’s audio isn’t being sent live. Can you confirm if the WebSocket logs in VAPI show any missing or delayed stream.send_audio or stream.receive_audio events? @Ahmad Raza

YES - There are critical missing events indicating complete audio stream failure: Missing stream.receive_audio events: Deepgram WebSocket closed at 04:05:33.365 without capturing ANY user audio. There should be continuous receive_audio events from the user's microphone, but there are ZERO - proving no audio data ever flowed from user → Deepgram. Deepgram stream cleanup triggered immediately: The "Deepgram stream cleanup" and "Deepgram WebSocket cleaning up" logs at 04:05:33.365 indicate the receive stream terminated before processing any audio packets. One-way audio only: The AI successfully sent audio (ElevenLabs generated speech), but the return path (user audio → Deepgram) never initialized properly. The user couldn't be heard despite speaking. No transcript activity: Zero user transcripts in the entire call proves stream.receive_audio was completely absent - Deepgram never received audio packets to transcribe. WebSocket closure logs show the bidirectional audio streams died instantly, with the receive path (user→system) failing completely while send path (system→user) worked briefly before also failing.Retry

Why am I being pinged? I'm not on the Vapi team 😅