We noticed a significant increase in Fish Audio TTS timeouts when using LiveKit Inference. We had 14 timeouts on 10th August, which then increased dramatically to 205 timeouts on 14th August.
We’ve been using LiveKit Inference with Fish Audio (model: fishaudio/s2.1-pro) for our real-time voice agents since 2nd August earlier it was mix of (cartesia and inworld), and it had been very stable until these recent spikes. The only non-default parameter we’re passing is the volume setting; everything else uses the default configuration. (LiveKit Docs)
One example session is:
https://cloud.livekit.io/projects/p_3scf9tt1e2i/sessions/RM_7i5rmfKV9xM6/observability?mode=transcript
Some sessions experienced 2–6 TTS timeouts within a single conversation. Unfortunately, traces were not enabled for those sessions. However, here’s another session with logs that may help with investigation:
https://cloud.livekit.io/projects/p_3scf9tt1e2i/sessions/RM_sB6C7HumFo3G/observability?mode=logs
https://cloud.livekit.io/projects/p_3scf9tt1e2i/sessions/RM_pnv4as8CyUeb/observability?mode=logs&t=1786715485079
Could you please investigate whether there were any issues or regressions with the Fish Audio LiveKit Inference service around 10th–14th August? It appears something changed on the backend, as our application and configuration remained unchanged during this period.
"livekit-agents==1.6.10",
"livekit-api>=1.2.0",
"livekit-plugins-fishaudio==1.6.10"