French text to speech: where your text actually goes

When you hand an article to a speech tool, you are not handing it a command: you are handing it the text itself. For most content that has no consequence. For an embargoed piece, an internal memo, a document that is not public yet, the question becomes fair: where does this text go, who sees it, how long is it kept. Most product pages stay vague about it, not out of bad faith but because vague is easier. Here is how to read the path of a text through any vendor, and ours stated without rounding.

One generation is at least two steps

A speech tool almost never does everything itself. A generation breaks into two stages, and each involves a different party. First, text preparation: cleanup, extraction from a web page, sometimes translation or summarisation, image description. This is the stage that sees your whole article. Then the synthesis itself: the prepared text is passed to an engine that produces the audio file. These two stages can be run by two separate providers, in two separate countries, under two separate legal regimes. A vendor who talks about a single "in-house voice" and never names these building blocks is hiding, often without meaning to, half of the journey.

The questions to put to any vendor

Three questions are enough to see clearly. Where are the application data and the audio files hosted? Who runs each stage, and in which country, which decides whether or not there is a transfer outside the European Union? And finally, can this text be used to train models, yours or a sub-processor's? That last question is the most important and the least asked. Many synthesis engines reserve the right, by default in their terms, to reuse submitted content to improve their own models. That is not a scandal in itself, it is a common clause: but if your content is sensitive you need to know before, not after. A vendor who answers "it is secure" without naming its sub-processors has not answered the question.

Our path, without rounding

Here is ours, as it is written in plain terms in our privacy policy. The application data and audio files are hosted in the European Union, in Ireland, with the provider Supabase. The site and API are served by Vercel. Text preparation, the stage that sees your entire article, is run by Mistral AI, a company established in France, for every account: it therefore happens in Europe, with no transfer, and that provider's terms exclude using content submitted through its API to train its models.

Synthesis, by contrast, is run by Cartesia AI and by Google, companies established in the United States, depending on your account's audio tier. The prepared text is passed to them to produce the audio, and that is a transfer outside the European Union, governed by each provider's data protection addendum. We would rather write it clearly than let you guess. And on the awkward point: the terms of these synthesis providers allow them, by default, to use submitted content to improve their own models. So we cannot promise you, on our own, that this will never happen. If that use is incompatible with your requirements, you need to write to us: we then document the exact state of our settings with each provider and the guarantees that apply to your account. This honesty about what the tool does not guarantee is the same line argued in what French quality means for a voice tool: we say what we do, and what we do not.

Staying in Europe, and proving it

For organisations that cannot let a text leave Europe, there is a sovereign tier: there, a generation fails rather than routing outside the European Union. That is a deliberate choice, the opposite of a silent fallback. On the other tiers, if the European text-preparation provider is unavailable, the stage falls back to a backup provider in the United States so your generation still completes; but every audio produced keeps, in its provenance log, the location where it was actually processed. So you can check afterwards, article by article, where the text went. That traceability is what separates a promise from a guarantee, and it is the kind of requirement the French text to speech page prepares for.

One last point that often reassures publishers: the audio player embedded on our clients' sites collects no personal data on listeners. No cookie, no IP address written down, no tracker. One listen is counted per article, and nothing else. The data question does not stop at the text you send: it includes what the tool does, or does not do, to the people who listen. There too, the right answer fits in one verifiable sentence, not a general promise. As for retention: your content and your audio are kept for the duration of the contract, then deleted on request.

Read next

• Libraries and media centres: audio in the service of reading accessibility
• Integrating schema.org markup for a narrated article: using AudioObject well
• Audio, ADHD and attention: what the format actually helps with