Find where your voice agent's seconds go
The unit that matters in a voice agent is not the request. It is the turn: the span between the user finishing speaking and the user hearing anything. Generic request-latency tools cannot see it. This can.
1. Paste a turn log — it never leaves your machine
JSONL, one turn per line. Needs total_ms and a stages object; first_audio_ms and unaccounted_ms are used when present. Analysis runs in this page's JavaScript. No upload, no account, no telemetry.
2. Or run the profiler locally
git clone https://github.com/YOURNAME/voicebench && cd voicebench pip install -e . voicebench demo --turns 80 --out turns.jsonl voicebench report turns.jsonl
The recorder wraps each stage of your agent, so the numbers come from your real pipeline rather than a synthetic log. voicebench diff before.jsonl after.jsonl turns a fix into a number you can defend.
Not on PyPI yet — the name is already used by an unrelated package, so install from source until it is renamed.
The paid version: a turn-latency audit
Voice turn audit — $149
Send me a capture from your real agent (or a 20-minute screen-share). You get back:
- A stage-by-stage p50/p90/p95 breakdown of time-to-first-audio, with the single bottleneck quantified as a share of the turn.
- The unaccounted time — queue wait and thread scheduling, which is where most "the model is slow" investigations die.
- A ranked fix list with the expected ms saved per change, and what it will cost in quality.
- A before/after harness so the next regression is visible instead of felt.
Turnaround is same day. If there is nothing worth fixing I say so and you pay nothing.
Payment & contact: not configured yet. Reply to the listing you found this through, or set a real contact address in store/index.html before sharing the link.
Built by someone who got tired of averaging latency numbers that hid the one turn everyone noticed.