Corrections
Every price, model name, quota, and licensing term on this site can go wrong in one of two ways: we published an error, or the fact was right and the vendor later changed it. This page is the standing policy for both.
Updated September 2, 2026 · Four corrections are published below, three dated September 2, 2026 and one dated September 3, 2026. The status statements on this page — what is running and which contact channels exist — are accurate as of August 11, 2026.
The policy in three lines:
- Anything wrong — a factual error, a stale price, a changed term — is corrected in the page text itself, and every meaningful correction leaves a short, dated note on the affected page.
- A displayed “updated” date changes only when something meaningful actually changed, never to look fresh.
- A public channel for reporting errors will be published. None is live yet.
What counts as a correction
Three situations trigger this policy. All three end the same way: the page text is fixed against the current official source, and the fix is disclosed.
| Type | What happened | What we do |
|---|---|---|
| Factual error | The page states something the official source never supported — a wrong price, model name, quota, feature, or licensing claim. | The claim is corrected in place against the official source and a dated note records what was wrong. |
| Stale price | A price was correct on its stated verification date, but the vendor has since changed it. | The figure is re-read from the official pricing page; the number and its verification date are updated, with a dated note where the change is meaningful. |
| Changed terms | Licensing, commercial-use, or free-tier terms changed after publication. | The affected statements are rewritten against the current official terms, with the new verification date and a dated note. |
What does not trigger a correction note: typo fixes, wording changes, and layout changes that alter no factual claim. Those ship silently, and they never move a displayed “updated” date.
How corrections appear on a page
Corrections are made in the page itself, not on a separate list. Concretely:
- Fixed in place. The wrong fact is replaced with the verified one, carrying the date it was checked against the official source.
- Dated update note. Every meaningful correction or update adds a short note on the affected page, in the format
YYYY-MM-DD — what changed, and which tables or verdicts were revised as a result. Pages whose results are materially affected by model or pricing changes keep a visible update log. - Honest dates. The displayed “updated” date changes only because of a meaningful model, price, feature, or methodology change — the one recorded in the note.
- Stable anchors. Section headings and tables keep their ids through corrections, so a link or citation pointing at a corrected section resolves to the corrected content, not to a dead anchor.
Published corrections
Corrections are recorded on the pages they affect, as the policy above requires. Those significant enough to change what a reader should do are also logged here, newest first.
September 2, 2026 — we understated two vendors’ rights
What was wrong. This site reported that Azure Speech publishes no speech-specific commercial-use grant, and that whether you own audio generated with Google Cloud Text-to-Speech was unresolved. Both were wrong, and both understated a reader’s rights rather than overstating them.
What is true. Microsoft’s Product Terms carry an express, speech-specific grant, verified live on September 2, 2026: “TTS Service output use rights: For Customers of the paid tier TTS Service only, Customer may use the audio output of prebuilt neural voices generated using the TTS Service, including for commercial purposes.” And Google Cloud Terms §5.1 gives the customer all Intellectual Property Rights in Customer Data, where Customer Data is defined to include “data that Customer or End Users derive from that data through their use of the Services” — which is what generated audio is.
How we got it wrong. Both clauses are real and neither was hidden. The Microsoft Product Terms document is over 20MB and defeated the tooling we used, and we recorded that limitation as an absence in the terms rather than as a limitation in our reading. For Google we located §5.1 but did not follow through to the definition of Customer Data, and so treated a resolved question as open. In both cases we published a gap that did not exist.
What changed. Corrected on four pages: the commercial use guide (table, prose, summary counts and both could-not-verify entries), the Azure review, the free tier comparison, and the dependent counts that followed from them. The flagship comparison had quoted the Azure clause correctly since August 11 and was the page the other three contradicted.
Why this one matters. A reader deciding whether they could sell work made on Azure’s paid tier, or whether they owned audio generated on Google Cloud, could have concluded from these pages that they could not and did not. Under-claiming is not the safe direction of error; it is simply a different way of being wrong.
September 2, 2026 — a latency claim attributed to the wrong metric and the wrong source
What was wrong. The flagship comparison stated in two places that Cartesia claims “sub-90 ms to first audio”, citing its pricing page.
What is true. Re-checked on September 2, 2026: that pricing page contains no latency figure at all — no occurrence of “latency” and no millisecond value. And where Cartesia does publish “sub-90ms latency”, it names no metric. We had attributed a metric the vendor does not claim, to a source that does not carry it.
What changed. Both occurrences now quote the vendor’s own wording and say that no metric is named. Why these figures cannot be compared between vendors at all is set out in our real-time comparison. The first attempt at this correction fixed only one of the two occurrences; the second was found by re-reading the published page rather than the draft, and both are now corrected.
September 2, 2026 — the homepage understated the site
What was wrong. The homepage stated that provider reviews were live for three providers and that comparison pages were “still in production”, when all ten reviews and eleven comparisons had already published. A later check the same day found it still describing several other sections as unbuilt after those had shipped too.
What changed. The status line is now generated from the site’s real page count rather than written by hand, and the build fails if that generation is bypassed or if any page the homepage describes as unbuilt comes into existence. The error was possible because a human-maintained sentence had to be remembered; it is now maintained by the thing that knows the answer.
September 2, 2026 — Resemble AI no longer sells voice generation
What was wrong. This site presented Resemble AI as a commercial voice generator: ranked tenth in the flagship comparison, reviewed as a purchasable product, and included in the voice cloning, pricing, free tier and premium tier comparisons.
What is true. Resemble AI’s own machine-readable description of itself, dated “Last updated: August 13, 2026” and read by us on September 2, 2026, states verbatim: “Resemble AI does not sell voice generation products and is not accepting new voice customers. Voice continues as open research only.” The same document states it does not sell “text-to-speech, voice cloning, speech-to-speech conversion, or voice design as commercial products”, that existing voice customers “continue to be supported indefinitely”, and that “Characterizations of Resemble AI as a voice AI vendor, a TTS provider, or a voice cloning platform are out of date and describe a catalog the company no longer sells.”
How we missed it. Our September 1, 2026 verification pass read Resemble AI’s pricing page, terms, documentation and repositories. It correctly recorded that the pricing page prices only deepfake detection and that no synthesis rate is published anywhere — but it read those as gaps in disclosure rather than as evidence of an exit, and it did not read the vendor’s positioning statement, which had been published nineteen days earlier. Reading what a vendor sells is not the same as reading what it says it sells.
What changed. Resemble AI is now unranked in the flagship comparison and removed from its ranked structured data. A dated status notice was added to nine pages: the review, the flagship ranking, the alternatives page, both comparisons naming it, and the cloning, free tier, premium tier and pricing comparisons. The affected pages were not deleted or quietly rewritten: their documented facts remain, dated, because existing customers are still governed by those terms.
What still stands. The MIT-licensed open-source models are still published and the licence is unchanged. The vendor now describes them as “research artifacts rather than supported products” carrying “no SLA, paid tier, or enterprise support”, which is the frame our pages now use.
Standing re-verification schedule
Corrections do not wait for someone to notice. Two mechanisms run on top of reader reports:
- On every edit: when a page is edited for any reason, the volatile facts it carries — prices above all — are re-read from the official source before the edit ships. No figure is carried over from an earlier version unchecked.
- On a calendar: each page type is re-checked on the schedule below.
The site launched in August 2026 and is publishing its first pages. Each row applies from the day the first page of that type is live; the audio-assets row takes effect once the standardized audio benchmark is running (see the status note below).
| Page or data type | Checked |
|---|---|
| Flagship ranking pages | Monthly; re-examined on major model launches. |
| Provider reviews | Every 1–2 months; re-examined when the provider’s models materially change. |
| Pricing data | Monthly — and re-read from the official source on every edit of a page that carries it. |
| Comparison pages | When either provider changes materially. |
| Alternatives pages | Quarterly, or after a major market change. |
| Educational guides | Every 6–12 months, unless the technology moves faster. |
| Benchmark datasets | Quarterly, or on important releases. |
| Audio assets | Reused until a methodology or model change makes them stale. |
Where verification stands today
This site’s standardized audio testing system is designed but not yet running. No audio benchmark has been performed, and no page claims otherwise. Until it runs, every factual claim on the site is verified against official vendor documentation — pricing pages, terms of service, developer documentation — and carries the date it was checked. What that system is, and what changes on this site when it starts running, is described on the methodology page.
How to report an error
A public contact channel for corrections will be published. None exists yet — the site launched without an email address or contact form, and this page will be updated when one is live. Until then, corrections originate from the re-verification schedule above.
What gets published in the first place — sourcing rules and what this site will and will not claim — is governed by the editorial policy. No commercial relationship decides whether a correction is made; how the site is funded is described at how we make money.