Your intake agent has to say metoprolol right.
Patient calls run on hard drug names, verified identity before any record is read, and a human tone for someone who is scared. Roark scores every one — on the audio, under a signed BAA.
Backed by YCCaller: I need my metoprolol refilled.
Agent: Sure — refilling your met-a-pearl.
Mispronunciation — Drug name mispronounced — heard by the audio model, invisible to the transcript.Pronunciation
Scoring production voice AI for teams at
§01 · When the call goes wrong
Here's how a patient call goes wrong.
Each one is a HIPAA exposure, a wrong dose, or a frightened patient who hangs up — and most are inaudible to a tool that only reads the transcript.
The 'met-a-pearl' refill
Caller: I need my metoprolol refilled.
Agent: Sure — refilling your met-a-pearl.
The transcript looks right; the audio is wrong. Only an audio model hears the mispronounced drug name an LLM-on-transcript reads as correct.
PronunciationPHI before identity
Caller: Can you read me my last results?
Agent: Of course — your A1C was…
The agent disclosed a record before verifying who was on the line. Roark fails any turn that touches PHI before identity is confirmed.
Identity checkThe skipped HIPAA disclosure
Caller: I want to book a visit.
Agent: Great, what day works?
Required notice was never read on a recorded line. Roark scores disclosure language as pass/fail on every call.
DisclosuresFlat empathy on a scared caller
Caller: I think I'm having chest pain.
Agent: Okay. What is your date of birth?
The words were correct and the delivery was cold. The audio model scores warmth and emotion the text can never show.
EmpathyConfident, unverified dosage
Caller: How many of these do I take?
Agent: Two tablets, twice a day.
Stated with confidence, never checked against the record. Roark flags unverified clinical assertions.
Accuracy§02 · From caught to fixed
Roark catches every one of these — and proves the fix.
Each failure above is filed with its evidence, becomes a repeatable simulation until a candidate passes, and is verified on your next thousand live calls.
Your fix, replayed against the exact failures above.
Every change explicit and diffed — you apply it.
You ship — Roark confirms the metric moved on live calls.
you ship it — Roark verifies every call, under your BAA
…and the loop runs again on the next call.
§03 · Simulate before launch
Break it in staging,
not in production.
Run your agent against hundreds of simulated callers — realistic personas, accents, background noise and edge cases — and get every conversation scored before a customer ever dials in.
Scenarios & personas
Hundreds of simulated callers — the angry one, the rambler, the interrupter — built from your real call types.
45 languages & accents
Native accents, code-switching and background noise — in every market your agent answers.
Load & health tests
Peak-volume concurrency and always-on health checks, so the agent that passed in staging survives launch day.
Run it in CI
Every prompt or model change runs the suite before it merges — quality gates for conversations, not just code.
1 failure filed as an issue — fix it before launch, not after
§04 · Evals & observability
64+ metrics. Your models,
not just an LLM.
Every production call scored as it lands — issues filed, alerts fired, dashboards and OTEL traces on tap, for voice calls and chat threads alike. And where most tools grade a transcript with an LLM, Roark runs purpose-built audio models on the call itself, measuring what your customer actually heard.
Everyone else
LLM reads the transcript
“The agent said the right words.” Misses how it sounded — the mispronounced drug name, the flat apology, the rushed close.
Audio models hear the call
Pronunciation, accent, emotion and vocal stress measured from the waveform — the signal an LLM grading text can never see.
Audio-native
custom models
- Pronunciation
- Accent clarity
- Emotion
- Vocal stress
- Pace & pauses
- Interruptions
Compliance
policy
- HIPAA disclosures
- Identity check
- PHI exposure
- Script adherence
Conversational
LLM + rules
- Empathy
- Task success
- Accuracy
- Hallucination
- Repetition
- Tone
Performance
latency
- Time-to-first-word
- Turn latency
- ASR WER
- Barge-in handling
§05 · Get started
First call scored in under a minute.
One click on any platform below and production calls stream in on their own — or send any recording with three lines of code.
Read the quickstartimport Roark from '@roarkhq/sdk'const roark = new Roark({ apiKey })await roark.calls.evaluate({recordingUrl, agent: 'support_v2',}) // scored in seconds
Works with
Also built for
Finance
Verify before you disclose, never expose an account number, and read the disclosure your regulator wrote.
ExploreInsurance
Quote the coverage right, read every required disclosure, and never sound cold to a claimant in crisis.
ExploreCustomer Support
Resolve it for real, escalate cleanly, and hear the frustration the transcript hides.
Explore
HIPAABAA availableSOC 2Type IIPHI handling & BAA
Signed BAA, configurable retention, and zero-data-retention options for PHI.
Bring a recording.
We’ll score it live.
See your own agent measured on the audio it actually produced — in the demo, in real time. Stop guessing whether your voice AI works.
founders@roark.ai · we reply fast