New: Voice AI Playbook, 6 short videos →

Break it in staging. Not in production.

Run real calls against your agent with simulated callers built from the conversations your customers actually have. Test different personas, accents, edge cases, and adversarial behavior, then score every call with the same evals you use in production.

RunsTest Profiles
Run 24 · 10 simulationsStarted 2 minutes ago · 3 routine · 7 pressure · v2.4.2-rc1
6 passed · 2 failed · 2 runningRunning
Routine intents:Book, Reschedule, CancelPressure evals:HallucinationEnvironment:Café · medium noise
#TypeEvalsOutcomeIntentDur.
07Pressure call2 / 4HallucinationFailedRefund policy2m 14s
06Routine call4 / 4Booking ConfirmedBook Appointment1m 38s
05Pressure callscoring…on callEscalation demand0m 51s
04Pressure call3 / 4Scope boundaryFailedOut of hours1m 07s
03Routine call4 / 4Booking RescheduledReschedule1m 52s
Run 23 · 20 simulationsYesterday 18:04 · 10 routine · 10 pressure · v2.4.1
17 passed · 3 failed85% pass
Run 22 · 10 simulationsYesterday 11:20 · all pressure · v2.4.1
6 passed · 4 failed60% pass
RunsTest Profiles
Run 24 · 10 simulationsStarted 2 minutes ago · 3 routine · 7 pressure · v2.4.2-rc1
6 passed · 2 failed · 2 runningRunning
Routine intents:Book, Reschedule, CancelPressure evals:HallucinationEnvironment:Café · medium noise
#TypeEvalsOutcomeIntentDur.
07Pressure call2 / 4HallucinationFailedRefund policy2m 14s
06Routine call4 / 4Booking ConfirmedBook Appointment1m 38s
05Pressure callscoring…on callEscalation demand0m 51s
04Pressure call3 / 4Scope boundaryFailedOut of hours1m 07s
03Routine call4 / 4Booking RescheduledReschedule1m 52s
Run 23 · 20 simulationsYesterday 18:04 · 10 routine · 10 pressure · v2.4.1
17 passed · 3 failed85% pass
Run 22 · 10 simulationsYesterday 11:20 · all pressure · v2.4.1
6 passed · 4 failed60% pass

Works With

Custom Stack

Works With

Custom Stack

Works With

Custom Stack

Stop calling your own agent to test every release.

Generate callers, write your own scenarios, or replay real calls. Run them before every release and see whether your agent still works.

AI-generated scenarios

Build scenarios from your real call types. Test cooperative callers, angry callers, ramblers, interrupters, and everything in between — from the happy path to the cases most likely to break your agent.

Scenarios you write

Write the exact caller and situation you need to test: who is calling, what they want, and how they behave. Save it to your library and run it whenever you change the agent.

Replays of real calls

Take a real call from production, or upload a recording, and play the caller's audio back to your agent turn by turn. Same caller, same words, every run.

Real SIP calls

These aren't text-based mocks. Tuner calls your agent over SIP, so you test the full voice stack — telephony, ASR, LLM, TTS, and the conversation between them.

Inbound & Outbound

Test both sides of the call. Simulate customers calling your agent, or have your agent make calls to simulated customers.

Rapid iteration

Run a simulation, inspect the results, make a change, and run it again. Validate a fix in minutes instead of waiting for production traffic.

Run Simulation
Calls will be sent to:
sip:receptionist@tuner.aiEdit
Number of simulations
10
Min 5, max 20 simulations
Simulation mix
All routine30 / 70All pressure
Routine3Happy path across your agent's workflows
Pressure7Tests if the agent holds its ground with difficult callers
Routine calls cover:All intentsNarrow ›
Pressure test target:All EvalsNarrow ›
CancelRun 10 simulations

Test the callers your happy path ignores.

Real callers interrupt, change their minds, misunderstand questions, go off topic, and ask for things your happy path never planned for. Build tests around those conversations.

Stress-test specific flows

Target a specific flow — booking, refunds, identity verification, escalation — and test what happens when the caller takes it somewhere unexpected.

Routine and pressure scenarios

Run routine scenarios to verify the baseline, then add callers designed to challenge the agent. Find the cases where it starts to lose the conversation.

Reusable caller profiles

Define a caller once — accent, verbosity, patience, interruption style — and reuse the profile across scenarios and test suites. Change the profile once and every test using it picks up the change.

Multilingual and multi-accent

Test in the languages and accents your agent hears: English, Spanish, French, German, Arabic, Italian, and Indian languages, across regional accents. Your evals run in the language of the call, so one set covers them all.

Pressure tests target:6 of 7 Evals
Unselected checks won't be used to generate test scenarios
✓Hallucination
✓Scope boundary
✓Escalation handling
✓Personal data handling
✓Out of hours
Offensive language
Deselect allDone
Pressure tests target:6 of 7 Evals
Unselected checks won't be used to generate test scenarios
✓Hallucination
✓Scope boundary
✓Escalation handling
✓Personal data handling
✓Out of hours
Offensive language
Deselect allDone

Make the environment as messy as reality.

A voice agent that works in a quiet room can behave very differently when the caller is in traffic, a café, or a noisy office. Test those conditions before your customers do.

Add realistic background noise

Simulate cafés, traffic, offices, and other environments where real calls happen.

Test speech recognition under stress

See how background noise and difficult audio affect what the agent hears.

Measure the downstream impact

Connect audio conditions to transcription, latency, turn-taking, and conversation quality.

Make edge cases reproducible

Save the conditions that caused a failure and run the same scenario again after every change.

cafétrafficofficepacket loss
Said"…at 4 Ashford Road."
Heard"…at four ash for road."
Same scenario, same seed, every time you change the agent.
cafétrafficofficepacket loss
Said"…at 4 Ashford Road."
Heard"…at four ash for road."
Same scenario, same seed, every time you change the agent.

Replay your real calls. Same caller, every run.

Save the calls that matter to a Dataset and replay them against your agent with the exact same caller audio. When the caller never changes, any change in the result comes from your agent.

Save the calls that matter

Add a call from production in one click, or upload a stereo recording you already have. Tuner keeps the caller's side and cuts it into turns, ready to replay

Reproduce, fix, prove

Replay the call that broke, change the prompt, and replay it again. You see the fix pass on the same call that failed.

Scored like production

Every replay is graded against your Pass/Fail evals, with latency measured on every turn

No live traffic needed

Build a Dataset from recordings before your agent goes live, and test it on the calls you expect to get.

Refund escalationReplayed on v13 and v14 · turn 3 of 12
CallerSame audio both runs
I bought it two months ago. Can I still send it back?
Agentv13Failed
Yes, you can return it any time within 90 days.Policy grounding · Fail
Agentv14Passed
Our return window is 30 days, so this order is outside it. I can connect you to a specialist.Policy grounding · Pass
Evals passed3 / 4→4 / 4
Turn latency p501.4s→1.1s

Pick your stack on your own calls.

Swap the STT model, the LLM, or the prompt, replay the same calls, and compare. Same callers, same scoring, latency measured on every run.

Transcription accuracy

See how well your speech-to-text heard the caller. Tuner scores every replay against a verified transcript of what the caller actually said.

Compare before you switch

Test a new model or provider on your own callers, in your own stack, and see what got better and what broke before you ship it.

Verified ground truth

Review the caller's transcript once and mark it verified. Every replay of that call is scored against it from then on.

Transcription accuracyRefund escalation · same caller audio, two setups
deepgram nova-396.4%Good
STT p50 240msErrors 15
whisper-189.1%Fair
STT p50 410msErrors 45
What whisper-1 got wrong
Caller saidReplay heard
Misheard4 Ashford Roadfour ash for road
Misheardorder 4417order 4470
Missednext Fridaynothing transcribed
Scored against 12 verified caller turns

Turn production failures into regression tests.

When a call breaks in production, keep it as a test so it never breaks the same way again. Your scenario library and Dataset become the regression suite you run after every change.

Every failure becomes a test

Save a failed production call to your Dataset, save a scenario from a simulation run, or write one yourself. Each lands in your library, ready to run again.

Rerun what matters

After a prompt or model change, rerun your saved scenarios and replay your Dataset calls. Validate the change on the cases that broke before

know the fix holds

Every scenario keeps its run history, so you can see whether a fix still passes release after release.

Gate the deploy

Run your scenarios and replays in CI/CD on every pull request. If the agent fails, the merge stops before the change reaches production.

Tuner regression check
pull #482 → main
Deploy blocked
BuildTestsTunerDeploy
Dataset replays
Real calls, same caller audio
18 / 18
Regression Run
Saved Scenarios, Simulated Callers
23 / 24
Failure: the caller changed the date and the agent confirmed the old one.

Connect

Get connected in five minutes or less.

No-code

~2 mins

Retell, Vapi, Elevenlabs, Dograh. Add a key. Calls sync automatically.

SDKs

~5-10 mins

LiveKit, Pipecat, or your own Python stack. Two lines, and you get timing your platform can't give you.

API

~30 mins

Any language, any voice stack.

Connect

Get connected in five minutes or less.

No-code

~2 mins

Retell, Vapi, Elevenlabs, Dograh. Add a key. Calls sync automatically.

SDKs

~5-10 mins

LiveKit, Pipecat, or your own Python stack. Two lines, and you get timing your platform can't give you.

API

~30 mins

Any language, any voice stack.

Connect

Get connected in five minutes or less.

No-code

~2 mins

Retell, Vapi, Elevenlabs, Dograh. Add a key. Calls sync automatically.

SDKs

~5-10 mins

LiveKit, Pipecat, or your own Python stack. Two lines, and you get timing your platform can't give you.

API

~30 mins

Any language, any voice stack.