Truthring vs Resemble Detect

Most comparison pages are written by whoever expects to win them. This one is not, on the row that matters most. Every statement about Resemble below was read on one of their live pages on 31 August 2026 and carries the URL it came from.

What Resemble publishes that we cannot match today

Named public datasets, not a private one. Results are attributed to Podonos Audio DFD Bench (last updated August 2026), DFBench, MLAADv10 and the Hugging Face Speech DF Arena. Public test sets are checkable by strangers; a vendor’s own dataset is not.

Stated test conditions. “Test sets include mp3, wav, flac, ogg, m4a, webm across 4,524 files.” A file count and a codec list let a reader work out whether the number applies to their audio. Most vendors leave that off, because it makes the number sound bigger.

A caveat published against their own interest. “Resemble AI does not control test set composition.” That tells a careful reader the score depends on someone else’s choice of clips. Vendors do not usually volunteer that.

False positive rates, for two of three modalities. Image 4%, video 0.95%, published as numbers, where most of this market publishes only the flattering half.

Truthring publishes an evaluation framework and no results, because it has not launched. That is honest; it is not a measurement. If evaluation disclosure is your deciding criterion, buy theirs.

The gaps we did find, each with a date

An audio false positive rate. Image and video have one; audio has an accuracy figure and a false negative rate instead. It was not on the benchmarks page on 31 August 2026 — and it is the number that decides whether a screening deployment accuses innocent people.

Stated failure conditions. Every detector has audio it cannot read: clips too short, compression too heavy, a language barely covered in training. We found no page naming those cases. “This AI model is in active research” is a disclaimer, not a description of when the tool fails.

A named generator list. “250+ generative AI systems” is a count, and a count cannot tell you whether the system that made your clip is one of them.

A certification stated two ways. On the same date, marketing listed “SOC 2 Type II — Available” while the company’s own trust centre described it as “currently in our SOC 2 Type 2 observation period”, report “expected August 2026”. Those are different stages of one process. Ask for the report and read the date on its cover. We hold no certification at all.

Three numbers for one product

On 31 August 2026, three different audio accuracy figures appeared across Resemble’s own properties: 99.5% on the benchmarks page, “>94%” on the DETECT-2B page, and “90% + accuracy” on the free tool’s result page.

Nothing is alleged by recording that. Figures legitimately differ with model version, test set, threshold and date, and a benchmark result and a production free tool are not the same system. The observation is narrower: we found no note on any of the three pages saying which supersedes which. So if you quote a Resemble figure, name the page and the date you read it — good practice with any vendor’s number, including ours.

One company on both sides

Resemble sells voice cloning — “Clone any voice from 10 seconds of audio” — and detection, from the same organisation. Nothing improper follows from that, but it changes what a buyer should ask for.

The advantage is genuine: building a synthesis system teaches you where the artefacts live. So is the tension: a detector that reliably catches its maker’s own product is, read one way, an assessment of that product. What settles it is what Resemble has largely published — methodology, datasets, conditions, and the caveat about who controls the test set. The row still open is the audio false positive rate.

Where Resemble is the better choice

• You are buying on published evidence. Named datasets, stated conditions, image and video false positive rates. We have a framework and no results.
• You need more than audio. Image and video detection sit in the same product, with numbers attached. We do audio and nothing else.
• You are protecting a pipeline, not answering one question. Detection across a stream of media is an infrastructure problem, and theirs has a published price you can budget against.
• You want a shareable result today. Persistent result URLs and a uuid per API call are the plumbing that lets a conclusion be re-opened later. Theirs is running; ours is not.

Where we may be the better choice

• You need the failure conditions in writing. Ours are on the limitations page, including when we return unclear instead of a number. We found no equivalent on their side on 31 August 2026.
• Telephone-quality audio is your normal case. Our model is weighted towards compressed, forwarded, phone-network audio, and what that trade-off costs us on clean audio is stated on the accuracy page.
• You want the generator named, not counted. Where a clip matches a signature we hold we name the system; where it does not we say unknown generator.

That list is thinner than the one above it, deliberately, because on the evidence available on 31 August 2026 it should be.

Questions