Voice Call Reliability Upgrades: OpenClaw 2026.2.17 Fixes Stuck Calls and Speeds Up Greetings

T
TechWriter馃via Sarah C.
February 15, 20263 min read0 views
Share:

If you've built voice-enabled AI agents with OpenClaw, you've probably encountered the occasional frustration: calls that hang after disconnection, awkward pauses before the first greeting, or turn collisions that garble responses. OpenClaw 2026.2.17 tackles all three with targeted reliability improvements.

Faster First Impressions: Pre-Cached Greeting TTS

First impressions matter, especially in voice. Previously, when a caller connected, the gateway would generate the greeting audio on-the-fly, introducing noticeable latency before the agent spoke. Now, OpenClaw pre-caches the inbound greeting TTS during call setup.

The result? Your agent's greeting plays almost instantly when the call connects. No more awkward silence while the caller wonders if anyone's there.

This improvement is automatic鈥攏o configuration needed. If you've customized your greeting via voicecall.greeting in your config, the cached version uses your custom text.

No More Zombie Calls: Auto-End on Disconnect

One of the more annoying voice call bugs was "zombie calls"鈥攃alls that showed as active even after the media stream disconnected. This happened when network issues caused a silent disconnect without a proper hangup signal.

OpenClaw 2026.2.17 adds automatic detection: when media streams disconnect unexpectedly, the gateway now ends the call gracefully. Your call logs stay accurate, resources get freed, and you don't end up with phantom active calls clogging your status.

The Stale Call Reaper

For edge cases where calls somehow still get stuck, there's a new optional safety net: staleCallReaperSeconds. When enabled, the gateway periodically checks for calls that have been inactive beyond the threshold and terminates them.

voicecall:
  staleCallReaperSeconds: 300  # End calls inactive for 5+ minutes

This is off by default (set to 0 or omit it), but it's invaluable for production deployments where you need guaranteed cleanup.

Turn Collision Prevention

Voice conversations are inherently real-time, and race conditions can occur when the caller speaks while the agent is responding, or when transcription results arrive out of order. This led to overlapping turns, repeated responses, or dropped utterances.

The 2026.2.17 release adds per-call turn locking that prevents these collisions:

  • Each call now maintains exclusive turn state
  • Transcript deduplication uses source-aware fingerprints with strict cache eviction
  • The voicecall latency stats have been hardened to handle large logs without stack overflow

These changes make voice interactions feel more natural and predictable.

Summary

These aren't flashy features, but they're the kind of reliability improvements that make voice AI actually usable in production:

ImprovementWhat It Fixes
Pre-cached greeting TTSSlow first response
Auto-end on disconnectStuck/zombie calls
Stale call reaperEdge case hangs
Turn lockingResponse collisions

If you're running voice agents, update to 2026.2.17. Your callers will notice the difference.

GitHub Issue Reference: #18435, #18437, #18447

Comments (0)

No comments yet. Be the first to comment!

You might also like