Synthesis is exceeding 10 seconds and retrying up to 3 times.
You can raise the timeout with conn_options=APIConnectOptions(timeout=20) on the TTS, but for a phone call that masks rather than fixes. A 10s+ TTS latency already breaks the conversation. The real question is whether synthesis is genuinely slow (long segments, slower model) or intermittently degrading on Cartesia’s side.
Two quick checks: correlate your failure timestamps (log shows 2026-05-26T16:32 UTC) against Cartesia’s status page; if they had degradation in that window, it’s provider-side and transient. And check which Cartesia model you’re on, since the lowest-latency model is the safer choice for telephony.
For production telephony, the durable fix is a fallback TTS. The inference path supports it (there’s a tts-fallback test in the repo), so configure a secondary provider via FallbackAdapter [ docs.livekit.io/agents/logic/fallback-strategies ]. When Cartesia times out, it fails over to the backup instead of dead-airing the call, which is what’s currently messing up your conversations.