Skip to main content

Resource

Voice measurement benchmark methodology and limits

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.

We do not claim universal, out-of-the-box accuracy across arbitrary networks. Telephony codecs, room acoustics and microphone profiles shape the signal.

What is a physics-based measurement?

Sonotheia measures physical properties of the voice and the channel in call audio and reports each one as nominal, deviated or unavailable, with the validity of the recording. A reviewer decides what to do with the result. Sonotheia does not decide whether a voice is synthetic.

What was benchmarked?

Sonotheia's measurements are benchmarked on simulated wideband, G.711 and AMR-NB chains built from public spoofing corpora. They have not been validated on live telephone traffic. A real trunk adds handset, network and transcoding effects the simulation did not include, so it stays pending until measured in place. Zoom, Teams and WebRTC meeting audio are out of scope.

What does this leave unproven?

Simulated corpus transforms do not validate live telephone traffic or a particular customer trunk. No real-time, engine-coverage or detection-accuracy claim follows from them.

Does a benchmark establish compliance?

No. A technical test does not establish that an output satisfies a legal or supervisory requirement. Legal and compliance reviewers assess applicability and adequacy.

Vendor comparison at a glance

Evaluation dimensionSimulated codec benchmarkActual deployment
AudioPublic spoofing corpora with simulated wideband, G.711 and AMR-NB transformsCustomer traffic not validated by that simulation
Additional effectsOnly the documented transformsHandset, network, transcoding and workflow effects need measurement
Meeting audioNo channel clearanceZoom, Teams and WebRTC remain out of scope
DisclosureConfiguration documented internallyNo external methodology delivery is promised

Frequently asked questions

What is Equal Error Rate (EER)?
In binary classification, EER is the operating point where false acceptance and false rejection rates are equal. It is a benchmark metric, not evidence of performance on an untested channel.
Which codec transforms were tested?
Sonotheia's measurements are benchmarked on simulated wideband, G.711 and AMR-NB chains built from public spoofing corpora. They have not been validated on live telephone traffic. A real trunk adds handset, network and transcoding effects the simulation did not include, so it stays pending until measured in place. Zoom, Teams and WebRTC meeting audio are out of scope.
Does calibration eliminate false positives?
No false-positive reduction or elimination is claimed. A channel-specific evaluation needs explicit acceptance criteria and held-out evidence.

Official regulatory references

Ready to evaluate Sonotheia for your voice channel?

Request a demo