Resource
How to evaluate voice deepfake detection vendors
Evaluate voice-analysis vendors against the decision you need to support, the measured channel, the quality of their evidence and their deployment data terms. Distinguish a detector's classification claim from a measurement instrument and from speaker authentication.
Sonotheia offers bounded voice and channel measurements for human review. It does not claim a synthetic-voice determination, liveness verification or regulatory approval.
Start with the decision
Specify whether you need a physical observation, a synthetic-voice classification, speaker authentication or transaction authorization. Evidence for one does not establish the others.
Compare like with like
Ask what was tested, on which held-out population and codec path, and what the output means. A vendor demo and a corpus result do not establish performance on your calls.
Define an evaluation
Set acceptance criteria for the actual workflow, including false assurance, review burden, unavailable readings and data handling. Ask which results are demonstrated, proposed or still unmeasured.
Check the evidence behind promises
Treat missing methodology and unexplained outputs as open evaluation questions. Counsel should assess legal requirements; a score alone cannot establish regulatory compliance.
When is Sonotheia a fit?
Consider Sonotheia when a reviewer needs bounded physical observations beside their own judgment. Scope the workflow and evidence requirements before an integration commitment.
Vendor comparison at a glance
| Evaluation dimension | Sonotheia scope |
|---|---|
| Voice and channel measurement | Named physical observations, reported as nominal, deviated or unavailable |
| Human decision | Step up, escalate or continue remains with the reviewer |
| Data handling terms | Confirm for deployment: storage, retention, access and subprocessors |
| Benchmark disclosure | The benchmark configuration is documented internally, with the corpus, partition and simulated codec chain identified. No headline performance figure is published here. A corpus benchmark does not establish performance on a customer's calls. |
| Evidence record | A structured measurement record with the instrument version can support a compliance team's own review and SAR narrative preparation. It is not a qualified SAR filing or a determination that a reporting requirement has been met. The institution and its legal reviewers decide what to file. |
| Deployment | Scope the customer environment and acceptance criteria before use |
Frequently asked questions
- Are detection, authentication and measurement interchangeable?
- No. Evaluate each against its own intended output and evidence. Sonotheia does not supply a synthetic-voice or identity determination.
- Should we compare vendors on a common test?
- Use held-out audio representative of the intended channel and workflow. Measure operational review burden and false assurance as well as any classification metric.
- Is a public corpus enough for procurement?
- No. A corpus benchmark does not establish performance on your actual channel or workflow.
- What questions should legal ask?
- Review data retention, subprocessors, biometric template creation, cross-border transfer, permissible uses and regulatory representations.
- How long should a pilot run?
- Long enough to cover your intended workflows (wires, callbacks, resets); the length and acceptance criteria are set with you.
- Where can we see Sonotheia?
- Request a demo at sonotheia.ai/request-demo or explore the controlled demonstration at shema.sonotheia.ai.
Ready to evaluate Sonotheia for your voice channel?
Request a demo