AI Voice Detection Privacy

What leaves the browser

When you choose a supported audio or video file, the browser decodes the selected speech and prepares a mono PCM WAV sample at no more than 16 kHz. The analysis control limits the selected window to a maximum of 30 seconds. This reduces upload size and sends only the analysis sample rather than requiring the detector to process an entire long recording.

The prepared WAV is encoded for the detection request and sent to the site's server. The server validates the format and size, trims the sample again to the selected limit, and rejects unsupported or oversized input. Audio detection cannot happen entirely offline because the live model is reached through the server.

Third-party model processing

Online detection is provided through Modulate's Velma-2 Synthetic Voice Detection API. The short audio sample is sent to Modulate for analysis, so Modulate's data-processing terms may apply in addition to this site's privacy policy. If the result identifies the provider as Mock, the app is in demonstration mode and Velma-2 did not produce that result.

The site does not intentionally save the uploaded audio file after the request completes. Detection results may be temporarily reused through a bounded server cache keyed by a file hash, which avoids repeated provider calls for the same sample. The cache stores the result rather than a copy of the audio file.

Audio you should not upload

Do not upload confidential meetings, protected health information, legal communications, private customer calls, security codes, intimate recordings, or children's voices unless you have a lawful basis and permission to process them. Trimming to the suspicious sentence can reduce exposure, but it does not remove the need for consent or organizational policy review.

If a sensitive case must be investigated, start with non-audio evidence: sender details, timestamps, account history, payment instructions, or a trusted callback. An employer, school, newsroom, or regulated organization should use its approved incident process and vendor review rather than treating a public web tool as a private forensic system.

Practical privacy controls

Use the shortest clean segment that answers the question. Remove unrelated conversation before upload, avoid including names or account numbers when they are not needed, and keep the original evidence in an approved location. Do not share the public result as proof about a person; it describes the selected sample only.

Signing in with Google raises the free daily quota and provides account identity information described in the privacy policy. Basic checks can be used without login. Before using the service for repeated organizational work, review the full privacy policy, the provider disclosure, and any retention or consent requirements that apply to your jurisdiction and role.

Privacy decision checklist

Related voice detection guides

AI Voice Detection Privacy FAQ