Checking for AI-generated speech

How to tell whether a recording contains speech made by an AI voice tool, and what the scores mean.

Some recordings may contain speech that was not produced by a real person. Voxint can score each speaker turn for how likely it is to be AI-generated (a "deepfake"), and show you the result as a simple risk level: low, medium, or high. You decide what to do with that information.

This feature is opt-in. It requires a separate GPU service and must be enabled in Settings before it does anything. If you did not install it, nothing in this guide applies and nothing changes about how your recordings are processed.

Related how-to guides: Settings and troubleshooting · Review and adjudicate. Technical details: gpu-contracts.md.

---

Turning it on

1. Make sure the synthdetect service is running. If you followed the setup guide, it is already part of your stack.

2. Go to Settings → Synthetic-speech detection.

3. Turn on Enabled. This is the master switch. With it off, the feature is completely dormant.

4. Optionally, turn on Autogenerate. With this on, every recording that finishes processing is scored automatically. With it off, you score recordings one at a time (see below).

If your installation uses multi-user accounts, only an admin can change these settings. A reviewer account can view scores but cannot turn the feature on or off.

Scoring a recording

Automatic scoring (autogenerate on)

When autogenerate is on, there is nothing to do. Each recording is scored as soon as it finishes the pipeline. The risk chip appears on the recording's detail page when scoring is complete (typically a few seconds to a minute, depending on length).

Manual scoring

When autogenerate is off, or for recordings that were processed before you enabled the feature:

1. Open the recording's detail page.
2. Find the Synthetic speech detection panel.
3. Click Score for synthetic speech. (In multi-user mode, only admins can trigger manual scoring.)

Scoring runs in the background. Reload the page after a moment to see the result.

Reading the scores

Each recording's detail page shows a Synthetic speech detection panel once scoring is complete. The panel shows:

• Mean risk: the average risk across all speaker turns. Shown as a colored chip (low, medium, or high).
• Max risk: the single highest-risk turn. Also shown as a chip.
• A link to the full report.

The full report page breaks the score down by individual speaker turn. Each turn shows its raw logit (a number from the model) and its calibrated risk percentage.

What the risk levels mean

Chip	What it suggests
Low (green)	The speech is consistent with a real human voice.
Medium (amber)	The model is uncertain. Could be real or synthetic.
High (red)	The speech is consistent with known AI voice generators.

These are the model's estimates, not certainties. A high-risk score does not prove the speech is fake, and a low-risk score does not guarantee it is real.

Known limitations

The report page includes a limitations section. The two most important ones:

• Some AI voice generators partially evade the detector. Chatterbox is the most notable: the model catches it less reliably than other generators. (#252)
• Recording conditions can shift scores. Certain real-world corpora produce different score distributions than others, which can inflate false positives in some cases. (#253)

These are documented findings, not bugs. They are why the scores are presented as risk levels for your judgment, not as verdicts.

Turning it off

Go to Settings → Synthetic-speech detection and turn off Enabled. No new recordings will be scored. Scores that were already produced remain visible on the recordings that have them.

Runs that were already queued when you turned it off may still complete. This is normal and not a sign that the setting did not take effect.