Earshot — Productivity skill for Claude Code

earshot

**A blind, reproducible benchmark for production phone voice agents.**

You have two phone numbers and no idea what is behind them. Earshot tells you which one you should ship — and shows its work.

Most voice-AI comparisons are vibes: someone calls each number twice, listens, and forms an impression. Earshot replaces the impression with a call battery that provokes specific failures, a waveform analyzer that measures what actually happened to within 10ms, and a scored report card that has to cite evidence for every number it gives.

  A  B+   86.4/100
  B  D    62.1/100

  System A wins by 24 points and the margin is almost entirely turn-taking.
  A stops talking 270ms after you interrupt; B keeps going for 2.1 seconds,
  which on a real line reads as the agent ignoring you.

---

Why this exists

Three things decide whether callers tolerate a voice agent, and none of them appear in a vendor demo:

1. **False barge-in.** Every vendor demos interrupting the agent. Almost nobody demos a cough, an "mm-hmm", or a spouse talking in the background — the things that make an agent stop dead ten times a call. Earshot measures the *false* stop rate separately from the real yield rate, because they are opposite defects and a system can be terrible at exactly one of them.

2. **p90 latency, not median.** A system averaging 600ms with occasional 2.5s stalls feels broken. A flat 900ms feels calm. Earshot reports the whole distribution and the interquartile spread, and weights the tail.

3. **Fluent confabulation.** A voice that sounds wonderful while inventing a price is the most expensive failure in the set. It gets flagged CRITICAL regardless of how good the call sounded.

What it measures

Numbers come from the waveform, not from a model:

Measured	How
Turn latency (median, p90, worst, IQR)	Gap from caller speech offset to agent speech onset
Barge-in stop latency	Agent speech offset minus caller interr