Roark
The only voice-QA vendor I found that names its own scoring models and publishes per-minute rates you can compute before you talk to sales. Buy it for the audio-native metrics and the config-as-code surface, and accept that a seven-person company is grading your production calls.
PRICING
| Pay as you go | $0 + usage | $50 free credit, no card. 1 project, 5 seats, 10 concurrent lines, 45-day trace retention. Simulation $0.15/min, metrics $0.04/metric/min |
| Team | $500/mo | The $500 is included usage, not a fee. 5 projects, 10 seats, 25 concurrent lines, 90-day retention, priority email plus Slack. Simulation $0.10/min, metrics $0.02/metric/min |
| Enterprise | From $4,000/mo committed | Unlimited projects and seats, 100 concurrent lines and up, custom retention, SSO/SAML with SCIM, custom RBAC, IP whitelisting, data residency, signed DPA, uptime SLA, invoicing and POs. Simulation from $0.05/min |
No per-seat and no per-metric licence. Live production calls you already run are billed as metric evaluation only, with no simulation fee, because Roark did not place the call. On calls Roark does place, telephony and speech-provider costs are passed through at cost. Transcription is free if you send your own transcript, otherwise $0.024/min on every plan. Extra concurrency is $100 per 10 lines per month. EUR figures are converted at roughly 0.87 and are indicative only.
checked 2026-09-01 · vendor pricing page
Element scores
Strengths
The audio-native scoring is the real differentiator and it is not marketing. Roark runs purpose-built models on the waveform and publishes the metric list: overtalk ratio, failed barge-in rate, incorrect agent interruption count, time-to-first-word, ASR word error rate, pronunciation. An LLM reading a transcript cannot produce those numbers. The developer surface is unusually complete for a company this size, with config-as-code YAML in git, a dry-run diff before apply, a CLI, Node and Python SDKs, an MCP server on the free tier and OpenTelemetry traces. Pricing is the other surprise: full per-minute rates published on the website with a downloadable PDF, in a category where Bluejay, Hamming, Cyara and Braintrust enterprise are all sales-led. SOC 2 Type II and a HIPAA BAA are available on the free tier rather than paywalled at Enterprise.
Honest dings
Roark grades other people's agents and publishes nothing about the accuracy of its own. Prism and Vibe are named but not documented, there is no versioning policy, and Roark was absent from the 2026 Testing the Testers benchmark that scored three of its competitors on evaluation accuracy against human ground truth. The company is roughly seven people founded in 2025 on about $2M, which is real vendor risk for something you gate a release on. Retention is 45 days at the bottom, there is no vector store or embedding surface, no A2A, and no multi-agent coordination. Two independent reviewers also flag that a team with no production traffic yet gets less out of a production-first tool, and that the public documentation outside the API reference is thinner than at established platforms.
Sources (12) — every claim traceable
Every audit lists the research it rests on — transparency and traceability are the product. Tools evolve: each audit is a snapshot of its audit date, and re-audits supersede older versions (kept below for reference).
- roark.ai/pricing — Official pricing, checked in full: three tiers with published per-minute rates, included usage versus fee explanation, concurrency and seat limits, 45/90-day retention, provider pass-through at cost, transcription at $0.024/min, and the statement that SOC 2 Type II and a HIPAA BAA are available on every plan including pay-as-you-go (accessed 2026-09-01)
- roark.ai — Official product overview: the catch-simulate-review-verify loop, 64+ metrics grouped into audio-native, conversational, compliance and performance, named customer logos, integration list (Vapi, Bland, Retell, LiveKit, Pipecat, ElevenLabs, Kore.ai, Google), and the Node SDK call-creation example (accessed 2026-09-01)
- docs.roark.ai/llms.txt — Official documentation index used as the capability inventory: observability, PII redaction, metric types and Studio, collectors, thresholds, datasets, simulation flows and personas, templates, run plans, schedules, WebRTC and chat simulations, config-as-code, tool-call testing, CLI, Node and Python SDKs, MCP server, Okta SSO with SCIM, Snowflake export, and the full REST API reference including knowledge bases and issues (accessed 2026-09-01)
- docs.roark.ai/documentation/metrics/system-metrics.… — Official system-metrics reference: the four metric definition types, the named models behind them (Roark Vibe, Roark Prism, Roark Interruptions, Roark Quality Analysis, plus Hume Expression Measurement for vocal emotion), and the full per-metric list with output types and scope (accessed 2026-09-01)
- docs.roark.ai/documentation/sdks/mcp-server.md — Official MCP server documentation: npx install for Claude Code, Cursor and VS Code, the two exposed tools (documentation search and sandboxed code execution against the TypeScript SDK), and worked example prompts (accessed 2026-09-01)
- docs.roark.ai/documentation/simulation-testing/over… — Official simulation architecture: plans and runs, agent targets and endpoints, Improv versus Scripted customer flows, personas, run templates (red teaming, multilingual, load testing, tool call accuracy), thresholds turning metrics into checks, and the note that the API still calls customer flows scenarios (accessed 2026-09-01)
- roark.ai/security — Official security posture: SOC 2 Type II complete, HIPAA BAA with zero-data-retention options, annual third-party penetration tests, ISO 27001 in progress, SSO/SAML via Okta and Google Workspace, RBAC with audit logs, configurable retention and deletion (accessed 2026-09-01)
- roark.ai/about — Official about page: founders James Zammit (ex-AngelList) and Daniel Gauci Mizzi (ex-Akiflow, iGaming), three further named engineers, fully remote across San Francisco, London and Malta, and the stated design premise that transcripts hide failures audible in the call (accessed 2026-09-01)
- ycombinator.com/companies/roark — Official YC profile: founded 2025, Winter 2025 batch, San Francisco, team size 3 at listing, founder backgrounds, Product Hunt launch August 2025 (accessed 2026-09-01)
- speechmatics.com/company/articles-and-news/de-risk-… — Independent 11-platform comparison, 2026: places Roark as production-first observability and replay, confirms 2025 founding by Zammit and Gauci with YC and F-Prime backing, and flags consumption pricing uncertainty, compliance documentation thinner than Coval or Hamming, and limited independent practitioner commentary (accessed 2026-09-01)
- contextqa.com/blog/ai-voice-agent-testing-tools — Independent: summarises the 2026 Testing the Testers study (21,600 human judgments across 45 simulations, evaluation accuracy validated against ground truth on 60 conversations) which scored Evalion, Coval and Cekura but not Roark, establishing that no independent accuracy figure exists for Roark's judges (accessed 2026-09-01)
- pitchbook.com/profiles/company/753018-13 — Independent funding record: $2M raised, investors F-Prime Capital, Liquid 2 Ventures, Massive Tech Ventures, NVO Capital and True Ventures (accessed 2026-09-01)