Truthring vs the ElevenLabs AI Speech Classifier

We sell a competing product, so nothing here is neutral. What we offer instead is traceability: every statement about ElevenLabs below was read on one of their live pages on 31 August 2026 and carries the URL it came from.

What the classifier gets right, said first

It states the scope of a negative result, in its own words. Asked “Does it detect audio from other AI voice tools?”, the answer published at elevenlabs.io/ai-speech-classifier on 31 August 2026 reads: “No. The classifier is built to detect speech generated with ElevenLabs. It is not a general-purpose detector for every AI voice provider.” Most detectors leave the meaning of a clean result to the reader’s imagination. This one closes it in two sentences.

It names a case where it stops working, and the case is its maker’s own current model. The same page states that the classifier “Does not reliably classify audio generated with the Eleven v3 model”. A vendor publishing the model its own detector struggles with is the rarest disclosure in this category.

It dates the edge of its provenance signal. “No ElevenLabs audio created prior to June 2026 carries a watermark.” A date is falsifiable, and it tells you exactly which body of audio the newer detector cannot help with.

The page also publishes “99% precision, 80% recall” for unmodified ElevenLabs audio. Recall of 80%, stated openly, is more candour than a headline “99% accurate” ever contains. Beside those numbers we found no false positive rate, no dataset, no sample size and no test date.

A generator checking its own output is doing provenance, not detection

Did ElevenLabs make this? On that question ElevenLabs is not guessing. They hold the model, can place a signal at generation and read it back, and may hold a record of what was produced. If a mark survives to the file you are holding, that is stronger evidence than any statistical reading of a waveform, ours included.

Did a machine make this? There a first-party tool is scoped out by its own FAQ, honestly. A clean result means not ours. It does not mean not synthetic. On screen the two look identical — a low number, a reassuring colour — and if the clip came from another vendor’s model, or an open-weights system someone ran on a laptop, “not ours” is entirely correct and entirely useless.

So this is a comparison of purposes, not of quality. Where the origin is known, the platform wins. Where it is unknown — nearly every recording anyone brings to a detector — its tool was never in the running, and says so itself.

The watermark boundary sharpens that. “No ElevenLabs audio created prior to June 2026 carries a watermark” means every ElevenLabs clip made before that date falls back on the tool the same documentation calls legacy. If the disputed voicemail on your desk is from last year, the stronger method has nothing to read. And one question worth putting to any watermark vendor, asked rather than asserted because we have not tested theirs: what survives re-encoding, a screen recording, or a trim to eight seconds?

The structure, described without accusation

A company that sells voice cloning and also grades its own output sits inside a real tension: a detector that catches its own product is, read one way, a report card on that product, and a vendor in that position has no commercial reason to become good at detecting a rival’s output.

Both halves point away from suspicion here. Nobody knows a system better than the people who built it, and ElevenLabs has already published what its scope is — so the disclosure that normally settles a structural question is already on their own website. We sell no synthesis, which is worth one thing only: no second product our verdicts could embarrass.

Where ElevenLabs is the better choice

• Your question is specifically about ElevenLabs. A licensing dispute, a terms complaint, an account investigation. They answer from provenance; we would be inferring.
• The audio was created on their platform after June 2026. The watermark detector then reads a mark placed at generation, which beats waveform analysis, ours included.
• They publish two things we do not yet. A retention period in time units — voice data not kept “longer than 3 years after your last interaction” — and a certification claim. We have neither, because we have not launched.
• Their speech engineering is not in question, and it would be silly to pretend otherwise.

Where we are the better choice

• The clip could have come from anywhere. A scam call, a leaked recording, a voice note forwarded four times. By its own published scope, the classifier is not built for that.
• The audio predates June 2026, or is not ElevenLabs at all. No watermark exists to read, so the question returns to inference — the only thing we do.
• The answer has to survive being questioned. A verdict from a party with no stake in the generator, carrying a reference code and an engine version, is a different object in a dispute.
• You want the failure modes written down, including when we return unclear: limitations.

Questions