What text-to-speech tools do with the text you paste
The text goes somewhere. Whether it is stored, logged, used for training or discarded is rarely stated anywhere you can find it — here is how to work it out, and what happens here.
The voice quality is what everyone compares. The thing almost nobody checks is where the text went, which is odd, because you supplied it and it is the only part of the transaction that was yours to begin with.
Most of the time it does not matter. Sometimes it matters a lot: an unpublished script, a client’s copy under NDA, a legal letter, internal training material, a medical or financial document someone wanted read aloud. Those get pasted into free web tools constantly, by people who would never email the same file to a stranger.
Where the text actually goes
Almost every online tool is a front end. Your text travels from your browser to their server, and usually onward to whichever synthesis service actually produces the audio. That is two parties minimum, each with its own retention policy, and the second one is frequently not named anywhere on the site.
At each hop the text can be discarded after producing the audio, cached so an identical request is cheaper next time, logged as part of ordinary request logging, or retained for training — and these are wildly different outcomes that all look identical from where you sit, because in every case you get an MP3 back in two seconds.
The four questions that separate them
Is it stored, and keyed to what? There is a real difference between an entry filed under a hash of the content and one filed under your account ID or IP address. The first is a cache. The second is a history.
Is it in the logs? Request logging is routine infrastructure, and text sent in a request body often ends up in it incidentally rather than by design. This is the most common way text is retained by tools that sincerely believe they don’t retain it.
Is it used for training? For a plain synthesis tool the answer is usually no, because your text isn’t useful for improving a voice model. Where it becomes live is with anything that also does cloning, conversational AI, or “AI features” beyond reading text aloud.
Who else receives it? If the tool is a wrapper — and most free ones are — their policy governs only the first hop. Ask what the second one does.
How to get an answer
The privacy policy is the place, but read it for the specific words rather than the reassuring ones. “We do not sell your data” answers a question you did not ask. What you want is a sentence about retention: how long, keyed to what, and who else it reaches.
If no such sentence exists, that is itself the answer. A tool that has thought carefully about retention says so, because it is a selling point. Silence generally means nobody decided, which in practice means the default — kept, in a log, indefinitely.
And a blunt rule that has never let anyone down: if you would not be comfortable with the text appearing in a support ticket at a company you have never heard of, do not paste it into a free tool. Not because anyone is malicious, but because the path it travels is longer than it looks.
What happens here
Stated exactly, and checkable against the privacy page:
Your text is sent to be synthesised and comes back as audio plus word timings. The result is cached, keyed by a hash of the request — the text and the voice settings together — so an identical request doesn’t pay for synthesis twice.
Being precise about that cache, because the honest version is more useful than the flattering one: the stored entry does contain your words, since it holds the subtitle files and subtitles are made of the words. What it does not contain is anything identifying. It is filed under the content, not under you, and there is no account, no session and no user ID for it to be filed under instead. Two people submitting the same text reach the same entry and nothing distinguishes them. It is content, not a record of a visitor.
Your IP address is used for two counters — a short-window rate limit and a daily character total — and nothing else. Both are numbers, not logs. The daily counter expires by itself after 36 hours.
Your text is not logged, not used for training, and not sold, because there is no mechanism here that would do any of those and no business model that would want to.
The part worth repeating
None of this is verifiable by you from the outside, here or anywhere else. That is the uncomfortable truth of the whole category: you are taking a written claim on trust, and a written claim is exactly what an untrustworthy tool would also provide.
What you can do is prefer tools that describe the mechanism rather than the sentiment. “We respect your privacy” is unfalsifiable. “Cached under a hash of the request, counters keyed to IP, expires in 36 hours” is a description specific enough to be wrong, which is the only kind of claim worth anything.