Check for deepfake audio

Audio deepfakes use generated or converted speech to misrepresent a speaker or what they said. Text-to-speech can produce spoken words from text; voice conversion can alter the apparent voice in existing speech. Zhang and colleagues describe these as distinct generation approaches in their research survey.

A suspicious recording can also involve ordinary editing, a misleading excerpt or an incorrect description of when it was made. Those are different questions from whether a voice was synthesized. NIST’s technical report separates content generation, provenance and the context in which material is presented.

Write down the concern precisely: “Was this voice generated?” is different from “Was this sentence cut from a longer conversation?” or “Did this message actually come from my colleague?” The answer may require different evidence for each question.

An impersonation request can be a form of social engineering even when the recording sounds convincing. If the message asks you to send money, disclose information or approve an action, use an independent way to check the request. The FTC’s family-emergency guidance recommends calling a number you already know belongs to the person. If they cannot be reached, another known family member or friend may help check the story.

Apply the same distinction to a work message: verify the claimed authorization through an established contact or approval route. A return number or link supplied in the suspicious message is part of what you are checking.

Keep a note of who confirmed what, through which contact and when. For example, confirming that a colleague did not authorize a payment helps resolve the immediate request. It does not identify the software that produced the recording or establish that every sound in the file is synthetic.

If you need to investigate the recording afterward, retain that separate question. A practical decision about a request and a technical conclusion about an audio file can have different supporting evidence.

Save the received attachment and retain the message or conversation around it. Record where it came from, when you received it, the claimed sender and any known forwarding history. Keep uncertain details marked as uncertain.

The file you received is not necessarily the original recording. SWGDE’s audio-authentication guidance distinguishes an original recording from a transcoded or re-encoded version. An exported or forwarded copy may contain useful sound while lacking information available in an earlier file.

Make a separate working copy before trimming, changing the volume, converting formats or applying cleanup. Keep a short record of each change and the tool used. Preserve the untouched received file alongside those derivatives rather than replacing it with the result you find easiest to hear.

If the earlier source remains available, ask for that file and an explanation of how it was obtained. Avoid creating another generation of the recording by playing it through speakers and recording the playback when a direct file copy is available.

SWGDE recommends protecting and, when possible, hashing the first removable instance of the original data. A file hash can help document whether particular bytes changed after the hash was recorded. It does not establish who spoke, whether the words are true or whether the file was already altered before you received it.

An unusual pause or a change in background sound can give you a timestamp to investigate. Treat these apparent red flags as observations, not as a universal test for generated speech. Record the observation separately from your interpretation: “a sudden change at 00:24” is more useful than treating the change itself as proof of a deepfake.

In a 2025 study, Sarah Barrington, Emily Cooper and Hany Farid asked 300 listeners to classify individual clips as real or AI-generated. Mean accuracy was 67.4% for genuine clips and 60.8% for the generated clips. Their separate identity task asked a different question about whether two voices belonged to the same person.

These are results from a controlled study using English-language recordings and a particular voice-cloning setup. They are not a universal detection rate for phone calls, every listener or every current generator. They do show why confidence in recognizing a voice should not replace an independent check.

The authors also found that some cues listeners reported using were not dependable general indicators. SWGDE treats critical listening as a way to guide further examination, with conclusions supported by a defined set of analyses rather than one isolated observation.

Before using a result, identify what the tool is designed to recognize. A classifier examines patterns in sound using its trained model. A watermark checker looks for a signal added through a supported generation process. These approaches examine different evidence, so their results must be interpreted within the tool’s documented scope.

A classifier can make errors in both directions. It may flag genuine speech as synthetic, or miss generated speech. NIST describes additional difficulties when a detector encounters an unfamiliar generation model, background noise, multiple speakers or a language different from its development data. Zhang and colleagues also discuss compression and real-world recording conditions as challenges.

Check the tool’s documentation for:

• The generation systems, models and languages it supports.
• Minimum duration, file requirements and the actual segment analyzed.
• Whether processing, editing or compression affects its reliability.
• What its score means and how the decision threshold was selected.
• Published evaluation conditions and the limits of those tests.

For a concrete example, ElevenLabs’ public AI Speech Classifier page says it uses only the first minute of a sample and does not reliably classify audio generated with Eleven v3. That first-party documentation describes the tool’s scope; it is not independent proof of its accuracy on your recording.

If you keep a detector result, retain the tool name, date, version when available, input file, checked interval and original report. A result for one interval should not be described as a verdict on a longer recording. When tools disagree, keep their reports and investigate the difference rather than averaging their scores into a new certainty.

Metadata can help explain a file’s format and reported history. Content Credentials can carry signed provenance information, while watermark systems look for signals associated with supported generation processes. These signals should be interpreted according to the system that created and checks them.

C2PA’s explanation describes Content Credentials as tamper-evident, signed records. It also explains that reviewing the full history of an ingredient requires access to that ingredient’s data. The presence of a credential is therefore a reason to inspect what was actually recorded and validated.

NIST cautions that technical transparency does not guarantee trustworthiness. A genuine recording can still be presented out of context. Likewise, absence of a watermark or credential is not enough to establish that audio is human-made: the relevant signal may never have been added, or may not be available in the received file.

Keep the source message, earlier copies and surrounding context even when a technical indicator appears reassuring. They address parts of the recording’s history that a single indicator may not cover.

For a consequential dispute, prepare a specific question for a qualified examiner. SWGDE’s version 1.3 guidance begins with clarifying what the submitter is trying to establish and assessing whether an authentication examination can answer it.

Provide the file as received, any available earlier source, relevant timestamps, known recording or export information and a list of processing already performed. Explain the concern without presenting a guessed conclusion as an established fact.

Ask what the proposed examination covers, what material is needed and what limitations may remain. A useful report should make its scope, methods, observations and reasoning understandable. The examiner may need additional files or recording-system information before reaching a conclusion.

Discuss file handling before sharing identifying or confidential material. Confirm the actual transfer, access and retention arrangements with the recipient rather than assuming that a public upload tool is appropriate for the recording.

Making speech easier to hear is a separate task from assessing its origin or integrity. Cleanup can produce a useful listening copy, but that improvement does not by itself establish authenticity. Keep the unprocessed file available for any examination.

This guide compares published forensic guidance, a controlled human-listening study, technical research and current first-party descriptions of provenance and detection tools. Sources were reviewed through October 31, 2026. The listening figures are reported study means, not new measurements by Media Medic. A study coauthor disclosed a role in a company developing detection tools.

No disputed recording was examined and no detector was tested for this guide. The 2018 SWGDE guidance supports examination planning and file handling; it does not validate a modern deepfake detector. Tool coverage can change, so check its current documentation before relying on a result. The guide does not determine a particular recording’s authenticity or its legal use.

Check an urgent request independently, preserve the recording and keep observations separate from conclusions. If a technical question remains important, seek an examination with a clear scope and stated limits.

Media Medic’s audio authentication guidance explains how that question differs from restoration. Ask about the specific concern, proposed deliverables, fee and timing before commissioning an assessment. A free restoration review should not be treated as a deepfake-authentication conclusion.