2026-03-03-fix-pipecat-audio-quality.md

Fix Pipecat Audio Quality Issues

Problem

Poor sound quality during AI calls traced to several bugs in pipecat-manager's media stream handling code (Go side).

Approach

Fix four issues in the Go audio pipeline:

1. Fatal error on audio write terminates output handler — A single transient write error to Asterisk kills all TTS audio for the remainder of the call. Fix: log and continue instead of returning.

2. Silent frame dropping with time.After GC pressure — pushFrame() uses time.After per frame (~50/sec), leaking timers until they fire. Frames are dropped silently with no monitoring. Fix: non-blocking fast path + reusable timer slow path + periodic drop logging via atomic counter on Session.

3. No frame drop monitoring — When audio frames are dropped due to channel backpressure, there is no logging or metrics. Fix: add atomic counter to Session, log periodically, and log total on session stop.

4. Pong routed through audio channel — Input receiver sends Pong via SendData which goes through the audio queue. In practice this is dead code (gorilla/websocket handles Ping/Pong internally via concurrent-safe WriteControl), but it is misleading and fragile. Fix: remove the SendData call and add a clarifying comment.

Files Changed

• models/pipecatcall/session.go — Add DroppedFrames atomic.Int64 field
• pkg/pipecatcallhandler/pipecatframe.go — Rewrite pushFrame with fast path and proper timer; add drop logging
• pkg/pipecatcallhandler/runner.go — Log+continue on audio error instead of return; clean up Pong handling
• pkg/pipecatcallhandler/session.go — Log total dropped frames on session stop

Resampling Investigation

Confirmed that no resampling occurs anywhere in the pipeline under normal operation. audio_out_sample_rate=16000 in Python ensures all TTS providers generate 16kHz natively. Go's GetDataSamples resampler is never invoked.