Inference latency has returned to normal.

A series of pathological calls significantly drove up inference latency for around an hour on Wednesday. Such calls can no longer be created.

Inference has been restored.

Inference is failing on voice calls.