Arize on debugging voice agents where the transcript looks fine but the call failed, and the tracing and evals you need for speech-to-speech systems.