subtopic

Recording, transcripts and latency metrics

Record calls (with disclosure) using egress to cloud storage; save the transcript and events per call; log metrics for each turn — end-of-turn delay, STT, LLM time to first token, TTS time to first audio — and compute voice-to-voice latency p50/p95. Then cut latency: a smaller/faster model, shorter prompts, region close to India, pre-warmed workers, filler phrases.