Use short, structured prompts, enable TTS chunking, set silenceTimeoutSeconds and maxDurationSeconds, implement model fallback on slow responses ( GPT-4o) and reduce stuttering or timeouts..