Benchmark your voice agent with real phone calls.
3 free checks
Add this line to your agent's greeting or prompt
check starting
this page updates live. keep the link to come back.
| Same model, three test paths | WER |
|---|---|
| Vendor's own test set | 5.26% |
| Curated public benchmark | 12.8% |
| Real phone audio | 25.2% |
| What you get | Measured how |
|---|---|
| Task success | Deterministic checks on facts the caller planted |
| Entity accuracy | Ground-truth readback over 8kHz telephony |
| Latency | P50 / P90 from the caller's ear |
| Turn-taking | Barge-in recovery, dead air, double-talk |
The rules
- Your number is verified before we dial.
- Reports are private by default.
- Recordings deleted after 30 days, or the moment you say so.
- Synthetic callers only. No customer data needed. Not for PHI.
- Unmeasured cells render null.