What this does, and what it won't promise.

Yes. There is no account, no paywall, no watermark and no trial. There is no paid tier to upgrade to, because there is no revenue model — the tool exists because it is useful.

No. There is no sign-up, and nothing to log in to. Your last-used voice is remembered in your own browser's local storage, which never leaves your machine.

SubRip (.srt) and WebVTT (.vtt), both UTF-8. You get them alongside the MP3 from the same synthesis, so they describe the exact audio you downloaded rather than a second, different rendering of the same text.

They are not estimated. The synthesis service reports a word boundary event for every spoken word, with an offset and duration in milliseconds, and the subtitles are built from those. Cues are then split at roughly 42 characters or five seconds, whichever comes first, so they stay readable.

Yes. The raw word boundaries arrive as bare tokens — "sentence" rather than "sentence," — so each token is located back in your original text and the cue is sliced out of the source with its punctuation intact.

5,000 characters per request. Longer text inside that limit is split, synthesised in pieces and stitched back into one file automatically — subtitles stay in sync across the join.

24 kHz / 48 kbps mono MP3. That is what the synthesis service is asked for, and the file you download is the exact bytes it returned — nothing is re-encoded.

Because the synthesis is not deterministic: the same text produces slightly different audio on each call. Streaming the audio and fetching timings separately would mean subtitles built against different audio from the file you keep. Subtitles are the point of this tool, so exact sync wins over a few hundred milliseconds of perceived latency.

There is no database and no account, so the text and audio you submit are never tied to you. Synthesised responses may be cached, keyed by a hash of the exact request, so that identical requests do not hit the upstream service twice. Cached entries are content, not identity — there is no record of who asked for them.

The site loads Google Tag Manager, gated behind Google's Consent Mode with every category — analytics and ads — set to denied by default. A banner asks first, and nothing in that category is allowed to fire until you accept; reject it and the default denial simply stays in place. Full detail, including which third parties may be involved, is on the /privacy page.

We can't tell you that you can. Neural voices are licensed by whoever trained them, and this tool is not in a position to grant rights it doesn't hold. If your use is commercial, check the licensing terms that apply to the voice you are using, or choose a provider that grants commercial rights to you explicitly and in writing. That is a genuine limitation, not a formality.

The catalogue covers roughly 150 languages and locales. It is fetched live rather than hard-coded, so the voice list on this site is whatever the service currently offers.

The interface is English only, and there are no plans to translate it. The output is a different matter — it covers the full voice catalogue, which is most of the world's widely spoken languages.