NVIQ / METHODOLOGY
How NVIQ evaluates voice systems.
The purpose is to evaluate whether a system preserves and correctly uses the information a human communicated — not simply whether a transcript reads well.
1. Capture and preserve evidence
Instrumented targets provide raw interaction evidence. NVIQ validates that input and preserves source events separately from its own derived observations. Missing critical evidence cannot be invented.
2. Apply common, versioned measures
Benchmark families include uncertainty and abstention fidelity, speech/state consistency, shared context, prior contamination, turn-taking, confidence channels and context transformation. Failure gates prevent an average score from concealing catastrophic behavior.
3. Compare only what is comparable
Model, API, provider version, adapter, test environment and supported input channels must be identified for every report. A missing or unsupported dimension is N/A / not comparable, not zero. A comparison must disclose meaningful architectural differences and limitations.
4. Review, correct and retain history
Before naming a provider, NVIQ requires factual-review outreach, applicable terms and publication review, response-status accuracy, a correction and rerun pathway, and preservation of prior report versions. Provider review does not imply endorsement. Comments are reproduced only with permission.