Voice AI: ten questions, answered

The questions people ask before they trust a number of ours, with the answer and the page that proves it. Nothing here is a promise about quality; every answer points at a status, a margin or a test you can read yourself.

1. What does it measure for a voice agent?

The signal, on both sides of the agent: what it hears (the caller’s recording, the telephone band, noise, clipping, gaps) and what it sends out (the loudness and true peak of a text-to-speech file, its noise floor, bandwidth, pitch and voice quality). Every measure comes from one register, /lenses, and carries a status, a margin and a line on what it does not say. The atlas puts all of them on the path from the microphone to the recogniser, and /voice counts them live.

2. Does it predict how many words a recogniser gets wrong?

Not that we claim. We tested one measure, on 10 October 2026: 624 recordings of 16 invented Dutch reports read by three synthetic voices in 13 conditions, including a simulated telephone band, transcribed by two Whisper models. The signal-to-noise measure did not predict errors in street, number and town (AUC 0.52 to 0.55; 0.50 is a coin flip). Real speakers, other recognisers and other languages are not tested. The test and its limits are on where we were wrong and in what AudioLab does not do.

3. Is my audio uploaded?

In the browser, no: the check, the outgoing check, the QC view and the labs measure on your device, and your audio is never uploaded. The API and MCP are different by nature: the audio reaches the EU-hosted service, is decoded and measured, and only numbers are kept. On a direct call it is discarded inside the request; a file sent through the signed upload route is removed within the hour. The exact words, with dates, are on /privacy.

4. What do a status and a margin mean?

A status is a promise about evidence, not about quality. Validated means tested against a standard or a known value, with the source named; a proxy approximates something and is information, never a pass or fail; beta has no ground-truth test yet and is shown nowhere in the voice checks; withdrawn failed its test and says so. A margin says how far the number can be off. A difference between two recordings that is smaller than the margin is not a real difference. Read a response walks through a real answer field by field.

5. How is a verdict made?

A rule is your limit on one measure. A value past the limit is Fail; a value past the limit but closer to it than the measure’s margin is Check, because the measure itself cannot tell; a value inside is Pass. A proxy is shown, not judged. Three faults fail a file on their own when you leave them switched on: clipping, gaps of digital silence and a level so low that nothing is there. There is no composite score. Every result names the rules version, the engine version and a hash of the limits it was judged with, so two verdicts can be compared later. How the verdict is made shows it on a real example.

6. Who sets the limits?

You do. The outgoing check offers published platform targets, such as EBU R 128 with the tolerance of the target check, and a starting range we measured on clean synthetic speech, labelled as a measured range and not as a norm. Edit any limit, switch a critical rule off, and download the set as gate.json for your pipeline. A limit we never published does not exist; the page says so where you would expect one.

7. Can I compare two versions, or follow a model over time?

Yes. The A/B tab of the QC view takes an earlier and a current file and reads each difference against the margin of its measure: inside the margin is “no clear change”, not an improvement. Signed in, you can save a run: the numbers, your limits, the verdict and a label, never the audio. A saved run serves as the earlier version in A/B and as a point in the trend view, where every panel measure draws a line from your oldest run to your newest.

8. How do I run the same checks in a pipeline?

With a free API key from /account: the same measures over the API and MCP, up to 20 recordings in one batch call, and measureStatus in the answer so your gate can read the status and the margin. The CI gate guide has the script that reads gate.json and shows what a failed check looks like. The hosted API has no webhooks; the gate runs where your build runs.

9. What is free, and what are minutes?

Everything in the browser is free and uses no calls. A free API key carries 1,000 calls and 2 GB a month; without a key there is a trial of 20 calls a month. A key you create for voice work counts measured minutes instead of calls: a free allowance each month plus a one-time start allowance, shown on /account with the date it expires, and the API answer carries voiceMinutes with what the file cost and what is left. Paid minute packages are in preparation and have no date; nothing is sold until they exist.

10. What does it not do, and what is only a demo?

It does not read the words, does not identify or count speakers, does not score naturalness, does not verify watermarks or provenance manifests, and does not measure latency. Where a page could show a number for one of these, it shows “not measured” instead. The live monitor runs as a demo in your browser: it names a gap, a clipping run, a level drop, dead air or hum as it happens, and those are signals, not verdicts. The full list is on what AudioLab does not do.