Project · Time a naive voice loop and write its latency budget

Build a small script that simulates (or, if you have API access, actually calls) an ASR step, an LLM step, and a TTS step, instrumented with the TurnTimer pattern from this module. Run it for at least 20 simulated turns, report mean and p95 per stage, and write a one-page latency budget document that allocates a 500ms target across the three stages with justification for each number.

• A runnable script with three timed stages (real or simulated) and a TurnTimer-style aggregator.
• A printed report of mean and p95 latency per stage across at least 20 turns.
• A one-page latency budget allocating a 500ms target across ASR, LLM, and TTS with a one-sentence justification per line.
• A short note identifying which stage has the worst p95-to-mean gap and one hypothesis for why.

• The script cleanly separates the three stages into distinct, independently timeable functions.
• Both mean and p95 are reported per stage, not just an overall average.
• The latency budget's three numbers sum to the stated target (or the writeup explicitly notes the overage and what would need to be cut).
• The tail-latency hypothesis names a concrete, plausible cause (provider cold start, network jitter, model load) rather than a vague 'sometimes it's slow.'