An audio provenance certificate: proving what produced the spoken version of an article

Editorial AI policies are now common. Some newsrooms allow a language model to tidy a text before it is read aloud; others forbid any generative tool on anything published under their name. Whichever line a newsroom draws, someone eventually asks the uncomfortable follow-up: a legal director, an ethics committee, a reader writing to the standards editor. What exactly produced this audio?

"Trust us" is not an answer an auditor accepts. A policy nobody can check is a statement of intent. That is why every audio version WeDispatch produces comes with a provenance certificate: a public page, readable without an account, that records what happened while the audio was being made. This article walks through what the page shows, what it hides on purpose, and the cases it does not cover.

It is written during production, not afterwards

The certificate is not a summary drafted once the audio exists. It is built from a log written while the audio is being produced, by the same code that prepares the text and sends it to a voice. Each step that changes something appends a line. If the customer's pronunciation lexicon replaced three proper names, that is logged. If text preparation had to fall back to processing outside Europe because of an outage, that is logged too.

That last case is the one that gives the page its value. A certificate that only ever reported good news would prove nothing, and anyone reading it would know. Fallbacks are shown, and they are highlighted in the list rather than buried in it.

What the page shows

The top block describes the version currently online:

• the version number and the production date;
• the duration and the language;
• the language model line. In charter mode it says, in plain words, that none was used and that the text read is the journalist's own. In standard mode it says a model prepared the text, with its tier and where it was processed;
• the voice, with its tier and processing location;
• the preparation mode, charter or standard.

Below that comes a list of what was done to the text, written as concrete lines: a Markdown link reduced to its label, pronunciations added from the lexicon, passages spoken differently from how they are displayed, quotes read by a second voice, spoken quote announcements that never appear in the on-screen text.

A final block counts, by category, the passages that the French reading rules converted: numbers, dates, amounts, acronyms. Listeners hear those conversions; readers still see the journalist's text. This is the approach described on our French text-to-speech page: fixed rules you can check offline, rather than an opaque rewrite.

What it leaves out, deliberately

A certificate attests; it does not republish. So the page shows no article text, no headline, no publisher name and no account identifier. You can forward it to an outside auditor without handing over the content itself.

It also does not name technical suppliers. The log records the tier and the location of each processing step, not whichever vendor was under contract that quarter. What a newsroom needs to demonstrate is where its text went and what kind of processing it went through. If you want the full route a text takes, where your text goes with French TTS covers it.

The page is also excluded from search engines. It is a document for a file, opened by invitation, not one more page competing in results.

Where to find it

In the customer dashboard, each audio's detail page has a Provenance block with a "See the public certificate" button, and the audio list offers the same link on every finished row. The address opens in a new tab and can be pasted into an email, an audit report or a reply to a reader as it is.

One version, one certificate

Every generation, translation or correction creates a numbered version. Nothing is edited in place. When a sentence is fixed after publication, the previous version stays exactly as it was, and the certificate describes whichever version is live. The correction leaves its own trace, so an audio produced in one pass can be told apart from one that was patched later.

Where it stops

Three situations produce no certificate:

• audio produced before the provenance log existed. The page says so rather than inventing content;
• private listening in WeDispatch Radio, such as personal documents or PDFs. Those are never published and have no public page at all;
• audio that is not finished yet. The button only appears once production is complete.

One more boundary worth stating: the certificate describes what our pipeline did. It knows nothing about what happened to a text before it reached us. If an article was drafted with a generative tool inside your CMS, nothing on our side can see that.

What it is actually for

For a newsroom that bans generative AI on its copy, the certificate of a charter-mode audio is the missing piece of evidence: it states that no language model touched the text that was read, and lists the only rules applied. How that mode works is explained in French TTS without generative AI.

For a publisher that does accept a preparation model, it answers a different question: where that model ran, and whether a given article, on a given day, was ever processed outside Europe. The neighbouring question, whether listeners must be told they are hearing a synthetic voice, is covered in should you disclose a synthetic voice.

The idea behind it is simple. A sales promise carries no weight with an auditor. A page that records what happened, including the times it went wrong, does.

Read next

• Security questionnaire for an audio tool: what to ask, and our answers, including the noes
• Breton, Basque, Corsican, Alsatian: why French text-to-speech trips on regional names
• A woman's or a man's voice for a quote: what French grammar reveals about who is speaking