Deepware checks faces, not voices
If the scanner is Deepware, no. Their own FAQ says so twice: “Truthring does not detect whether the voice is manipulated but only the face”, and “We are only dedicated to detecting AI-generated face manipulations.” Voice detection is described there as in development. Checked on deepware.ai, 31 August 2026.
So a clean Deepware result is a statement about the face on screen and nothing else. If the picture is genuine footage and the audio has been replaced, that is precisely the case a face scanner is supposed to pass — and it will.
This page used to be a head-to-head comparison. It is not one any more: a disclosure table pitting an audio detector against a tool that does not analyse audio would be meaningless and slightly dishonest.
How this page is sourced. Every statement about Deepware below was checked on deepware.ai on 31 August 2026. Several of their pages could not be reached by our fetcher and are marked as fetch failures, not as absences. Corrections are published with a date: corrections@aivoicedetctor.com
Why there is no comparison table here
Every other page in this section carries a two-column table of what each side publishes: false positive rate, methodology, failure conditions, retention. Setting that table against Deepware would produce a column of blanks, and a reader skimming it would think a vendor had failed to disclose things about a product that does not exist.
Deepware has not declined to publish its speech accuracy. It does not do speech. Presenting that as a gap on a scorecard would be the sort of quiet misrepresentation this whole site was built to argue against, so the table is gone and this explanation is in its place.
Faces and voices are different problems
A visual deepfake detector looks for things that only exist in pictures. The seam where a swapped face has been blended into a head. Lighting on a cheek that does not match the room. Eyes that blink at the wrong rate, a jaw that moves in a way the neck does not follow, inconsistency from one frame to the next. Almost all of it depends on there being frames, and on there being a face inside them — which is why Deepware requires at least one detectable face before it will scan at all.
Speech offers none of that. No face, no lighting, no motion, no temporal frame structure of the kind a video model expects. The evidence is elsewhere entirely: the fine structure of the spectrum in the highest bands, the way a vocoder reconstructs breath and sibilance, the statistics of pauses, the flatness of prosody across a long sentence, the small mouth and throat noises a person cannot help making.
The two are related the way fingerprint analysis and handwriting analysis are related. Both are forensic, both are pattern work, and expertise in one does not carry over by itself. A team excellent at video may also be excellent at audio, but only because they built a second thing. Deepware says plainly that they have not built it yet, which is more than most vendors manage about anything.
Why the soundtrack is the harder half to check anyway
Even where a video-first tool does look at audio, the audio it receives has taken a rougher path than the file you started with: muxed into a container, encoded with a codec chosen for the picture, resampled, sometimes loudness-normalised. Compression discards exactly the high-frequency detail speech detection leans on hardest, which is why the same voice is easier to judge as a standalone file than as the soundtrack of an upload. That is the practical reason to separate the two questions rather than hope one tool answers both.
Which question, and where to take it
| What you have | What you are actually asking | Where that gets answered |
|---|---|---|
| A video with a face you doubt | Has this face been swapped or generated? | A face-oriented scanner. Deepware is free, self-serve and built for exactly this |
| A voicemail, voice note or call recording No picture at all | Was this speech generated? | An audio detector. This is the only thing Truthring does |
| A video where the face looks fine but the voice sounds off | Two questions, not one | Both, run separately. Extract the original audio track without re-encoding it, then compare the two answers |
| A video, and you want one number for the whole thing | Is this file fake? | Nothing honestly gives you that. A blended score across a real face and a replaced voice is uninformative about both |
| A live call happening right now | Can I stop this before the money moves? | Neither. Both Deepware and Truthring work after the fact, on files already recorded |
Routing, not scoring. Nothing in this table ranks one tool against another; they answer different questions. Checked 31 August 2026.
The thing Deepware does better than anyone else we checked
They publish no accuracy percentage at all. Of the ten vendors we examined on 31 August 2026, Deepware is the only one claiming none. The homepage says “Accurate, reliable results” and attaches no number to it. The FAQ goes further and actively disclaims certainty: “Deepfakes are not a solved problem yet.”
We regard this as a credit rather than an omission, and it is not a close call. A vendor that refuses to quote a figure it cannot support is behaving better than one that quotes an unsourced 99%. An unsupported percentage is worse than silence, because a number invites reliance: somebody puts it in a report, a tribunal reads it as measured, and nobody can trace where it came from. Silence at least tells the reader they still have to think.
What this site asks for beyond honesty is a rate you can act on, published in both directions with the test set attached. Deepware does not offer that either. But the ordering matters: publishing no number is the second-best position, and publishing an invented one is the worst.
And the clearest limitations statement of anyone checked
Their documentation states hard operating limits rather than hedging: a maximum of 10 minutes per video, at least one detectable face required, 1920×1080 recommended, and a frame-rate note reading “1 FPS is a good compromise between scanning speed and detection accuracy”. That last sentence is a vendor telling you where they traded accuracy away and why. Almost nobody does that.
The scanner is free, self-serve, and works after the fact on uploaded files; the company describes itself as founded mid-2018. No individual is named — the site credits “the deepware AI team” — and we found no certifications and no published legal entity on the pages we could reach on 31 August 2026.
[VERIFY: fetch failed 2026-08-31, check manually before publishing] Deepware’s privacy policy, terms, scanner and result pages all timed out for our fetcher on three attempts. Retention, result metadata and reference identifiers are therefore unchecked, not absent. Nothing on this page asserts anything about them in either direction.
Where Deepware is the right tool and we are not
- There is a face on screen. Decisive. Most of the available evidence is visual, and we cannot see any of it.
- You want to check something at no cost, right now. Free public scanning reaches ordinary people that paid products never do, and we are not going to be sniffy about it.
- You are sorting a pile of media quickly. An upload that returns a fast read is the right shape for triage; a careful per-clip analysis is the wrong shape.
- You want to be told where a tool stops working. Ten minutes, a detectable face, a recommended resolution, an explicit accuracy-for-speed trade. That is better documentation of failure than most of this industry offers.
Where an audio detector is the right tool
- There was never any video. Most voice fraud involves no picture at all — a call, a voicemail, a message on WhatsApp. The audio is the entire body of evidence.
- The voice is the disputed part. Real footage with replaced speech is the exact case a face scanner is designed to pass.
- The result will be challenged. That is what a published error rate, a versioned method, a Ringmark reference and an explicit unclear verdict are for.
Truthring is pre-launch. The commitments above are design commitments published in advance of any measured result, and no figure on this site claims otherwise — see accuracy and limitations.
Questions
Will Deepware check the voice in my video?
No. Their own FAQ says it in two places: “We do not detect whether the voice is manipulated but only the face” and “We are only dedicated to detecting AI-generated face manipulations.” Checked on deepware.ai, 31 August 2026. Voice detection is described there as in development. If your doubt is about the speech, a face scanner will not answer it.
If Deepware clears my video, is the voice cleared too?
No, and this is the failure mode that matters most. A video can be genuine footage of a real person with the audio replaced. A face scanner looking at that video is looking at a real face, and a clean result is exactly what it should return — while saying nothing at all about the words. The two questions have to be asked separately.
Does Deepware publish an accuracy figure?
No, and of the ten vendors we checked on 31 August 2026 it is the only one claiming none. Their homepage says “Accurate, reliable results” without attaching a number, and the FAQ actively disclaims certainty: “Deepfakes are not a solved problem yet.” We think that is the more honest position, and better than an unsourced 99%.
Why can a detector not do faces and voices at once?
Some can, and the multi-modal platforms do. But it takes building two things, not one thing that generalises. Visual detection reads blending seams, lighting that does not match the room, and inconsistency between frames. Speech has no frames and no face; the evidence is in spectral fine structure, the way a vocoder reconstructs breath and sibilance, and the statistics of pauses. Expertise in one does not transfer by itself.
What should I do with a video whose voice sounds wrong?
Check the picture and the sound separately, and keep the original file. Extract the audio track without re-encoding it if you can, because a second round of compression removes the fine detail speech analysis depends on. Then run a face-oriented scanner on the video and an audio detector on the track, and treat disagreement between them as informative rather than as an error.
Does Truthring analyse video?
No. We take audio and nothing else. Hand us a video and we are analysing one track and blind to everything on screen, which is a limitation before it is a virtue. On a file with a face in it, most of the available evidence is visual and we cannot see any of it.
Sources
- deepware.ai — scope statements on face versus voice, absence of any accuracy percentage, the “Deepfakes are not a solved problem yet” disclaimer, operating limits, founding year, team attribution. Checked 31 August 2026.
- Deepware privacy policy, terms, scanner and result pages — could not be read on 31 August 2026 after three attempts. Marked for manual check. Nothing about retention or result metadata is asserted from them either way.
Claims about Deepware checked 31 August 2026 against publicly available material. Re-checked quarterly. Corrections are published with a date rather than made silently.