HomeVoice Detection › For Journalists

Verifying audio before you publish.

A clip lands in your inbox and the desk wants to know by four o'clock. Detection is the fastest check you can run and the weakest one to publish on — it speaks to how the audio was produced and says nothing about who is speaking, when, or in what context. This is where it fits in a verification workflow, and where it does not.

Check an audio clip free The workflow ↓
3 checks a day, no signup Audio never stored
Scope

What detection can and cannot establish.

Getting this boundary right is the whole job, because the failure mode here is public.

✓ It can indicate

  • Whether the speech carries the acoustic signature of a physical recording
  • That a clip is worth escalating for closer examination
  • Which of twenty submitted files to look at first

× It cannot establish

  • Who is speaking — that is speaker verification, a different task we do not perform
  • When or where a recording was made
  • Whether genuine audio was spliced, cut between sentences or misattributed
  • Whether the claim being made in the recording is true

That third exclusion deserves emphasis, because it is the one that burns newsrooms. A recording can be entirely authentic audio and entirely misleading — real sentences from a real occasion, assembled to imply something that was never said. A spoof detector will call that human, correctly, and uselessly. If your actual concern is manipulation rather than synthesis, you are holding the wrong instrument.

Not forensic-grade. A TextSight voice result is a triage signal, not proof, and it must never be the sole basis for a disciplinary, employment, financial or legal decision. Our full position on what a result is worth is here.
Workflow

Where it fits in a verification workflow.

Detection is step three, not step one. The cheap questions come first because they resolve more cases.

  1. Provenance before analysis. Who sent it, who had it before them, how many hands has it passed through? Trace it as close to origin as you can. Most files that fall apart, fall apart here — before anyone opens an audio tool.
  2. Preserve the original untouched. Keep the file exactly as received, with its container and metadata intact, and work only on duplicates. Note the hash and the date received. If this becomes contested, the untouched file is what matters.
  3. Run detection as triage. Extract clear continuous single-speaker speech, export as MP3, WAV, M4A, OGG or FLAC up to 10 MB, and check it. Record the conditions alongside the result — duration, format, how many hops.
  4. Run a control. Published, verifiable audio of the same speaker through a comparable channel. The comparison method is the single most useful thing available to you, because it partly cancels out the channel effects that otherwise dominate a lone score.
  5. Corroborate outside the audio. Does anyone place the speaker there? Does anything else from the same occasion exist? Does the content contain a checkable detail?
  6. Put it to the subject. Obvious, routinely skipped, frequently decisive — and required anyway before publication.
If the story turns on the recording itself, commission a forensic examiner. Chain of custody, container and encoding structure, edit-point analysis and an expert who can defend a methodology are a different discipline from an automated score. Budget for it when the stakes justify it — and if they do not justify it, ask whether they justify publishing.
Writing it up

How to describe a result in copy.

The language you use travels further than the finding does.

Detection results get flattened in transmission. Whatever hedge you write, a sub-editor may cut it and an aggregator will drop it. So the hedge has to survive being shortened.

AvoidPrefer
“AI analysis proved the recording was fake”“An automated detection tool assessed the audio as likely synthetic; the tool is not forensic analysis and its makers say it should not be relied on alone”
“99% certain it was AI-generated”Name the tool and describe the output in its own terms, with the audio conditions stated
“Experts confirmed the voice was cloned”Say who, with what method, and what they actually claimed
“The detector cleared the recording”“The tool found no evidence of synthesis” — which is not the same as authentication

Two further points worth carrying into copy. First, state the audio conditions: a result on a forwarded, compressed phone recording deserves a different weight from one on an original file, and readers cannot apply that discount if you do not tell them. Second, remember that false positives land on people who record well — professionally produced audio is mastered in ways that strip the untidiness detection looks for, so a polished recording is more exposed to a wrong flag, not less.

FAQ

Newsroom questions.

Can a voice detector tell me who is speaking?
No. That is speaker verification, a separate task, and our detector does not perform it. It estimates whether speech was machine-generated or physically recorded — not whose voice it is.
Is a detection result publishable evidence?
Not on its own. It is not forensic-grade and should never be the sole basis for a published claim that a recording is fabricated. Treat it as triage that tells you where to spend reporting effort, and if the story turns on the audio, commission a qualified forensic examiner.
What if the clip is a forwarded, compressed copy?
Expect meaningfully lower confidence. Compression removes much of the fine detail detection relies on, so an ambiguous result on a degraded file is usually a statement about the file rather than the voice. Chase the most original copy you can obtain before drawing conclusions.
Can it detect that genuine audio was edited or spliced?
No. Splicing is not synthesis — every fragment is authentic human audio, so a spoof detector will report human. Detecting manipulation requires discontinuity and provenance analysis, which is a different examination.
Does it work on non-English audio?
The analysis works on acoustic properties rather than on the language spoken, so it is not English-only. We have not published separately measured figures per language and will not claim parity we have not tested. Treat non-English results with the same caution, and use a same-language control clip where you can.
How should I store the file for a possible legal challenge?
Keep the original exactly as received, unmodified, with its metadata and container intact, and record when and from whom you received it. Do all trimming and converting on copies. An edited file is worth far less to an examiner than the file as it arrived.
Related

More voice detection guides.

One strand of verification. Never the story.

Three checks a day, free, no signup. Your audio is never stored.

Check an audio clip All voice guides
Triage signal · not forensic authentication · audio processed and discarded