Free
For your first agent. Find out what your calls actually score.
$0/mo
no card required
- 5,000 graded turns per month
- 2 scenario suites
- Accuracy, tone, and task-success scores
- 30-day score history
- 1 project, 3 teammates
Evaluation infrastructure for voice AI
Ship voice agents that do not break in production.
Sonora tests, monitors, and grades your voice agents on real conversations, catching regressions and bad calls before your customers do.
Free to start. No card.
suite: scheduling · deploy 4c81 · 128 calls
deploy gated
- 94
Password reset
8 turns clean
- 83
Plan upgrade question
interruption, 1.8s
- 61
Reschedule a cleaning
turn 6: task 41
failing turn · agent · turn 6
“Great, you are all set for Tuesday.”
Booked without reading back the date or time. Task success scored 41, against a suite baseline of 92.
alert sent to #voice-alerts, 41s after grading
Grades 500,000+ calls a month for 120 voice teams
- Voxly
- Dialpath
- CallOtter
- Meridian Air
- Fernwood Health
- Northbeam Legal
The problem
Voice agents fail silently.
One prompt change can break 1 in 5 calls. No exception is thrown, no dashboard turns red. The agent keeps talking, your customers keep hanging up, and the only way to catch it is listening to recordings by hand.
- Calls a single prompt change can break
- 1 in 5
- Errors surfaced in your logs
- 0
- Minutes to hand-review one call
- 11
128 production calls, one regressed. Nobody heard it happen.
How it works
From first call to gated deploys in an afternoon
- 01
Connect your agent
Point Sonora at your agent and import call transcripts and audio. Native adapters for Vapi, Retell, LiveKit, Pipecat, and Twilio, or push JSONL over the API.
- 02
Run scenario suites
Sonora replays scenario suites and scores every call, turn by turn, on accuracy, tone, and task success against your rubric.
- 03
Catch the regression
When a score drops below baseline, you get an alert with the exact turn that failed, before the deploy reaches customers.
Features
Everything between “it demos well” and “it holds up”
Scenario test suites for voice
Define the calls that matter, an angry cancellation, a rescheduling with a confirmation number, and rerun them on every change.
Automatic call scoring
Every call gets accuracy, tone, and task-success scores from a grader calibrated against 25,000 human-labeled turns.
Regression alerts on every deploy
Score drops against your baseline page you in Slack or PagerDuty and can gate the deploy in CI.
Turn-level failure pinpointing
Alerts link to the exact turn that failed, with the transcript, the audio, and the rubric line it violated.
Latency and interruption monitoring
Tracks time-to-first-word, dead air, and talk-over on every call, because a slow agent loses callers even when it is right.
Live demo
Watch Sonora grade a call
Three sample calls from a scheduling agent, graded with the real scoring rubric. Pick one and watch every turn get scored.
reschedule-a-cleaning.wav · scenario suite: scheduling
select a call to start
From the field
Teams stopped listening to recordings by hand
“We changed one line in the greeting prompt and Sonora flagged an 18-point task-success drop before the deploy went out. That single catch paid for the year.”
“We went from listening to 30 random recordings a week to reviewing 4 flagged turns a day. Same coverage, a tenth of the time.”
Pricing that scales with your call volume
Every plan grades real calls with the same scoring engine. Start free, upgrade when your agents do.
Billing period
Two months free when billed annually.Recommended
Pro
For teams shipping to real customers. Catch regressions on every deploy.
$249/mo
billed annually
- 100,000 graded turns per month
- Unlimited scenario suites
- Regression alerts in Slack, PagerDuty, and email
- Deploy gates for CI, block a bad prompt before rollout
- Turn-level failure traces with audio
- Latency and interruption monitoring
- 12-month history, CSV and API export
- Priority support
Scale
For platforms grading millions of calls with strict compliance needs.
Custom
- Unlimited graded turns
- SSO and SCIM
- On-prem or private-cloud deployment
- Custom scoring rubrics with a dedicated eval engineer
- 99.9% uptime SLA
- DPA, data residency, and custom retention
FAQ
Questions teams ask before connecting an agent
Your agent talks to customers all day. Hear about the bad calls first.
Connect an agent, run your first suite, and get a graded report on 5,000 turns this afternoon.
Free to start. No card.