Text to Speech Converter

Turn text into speech with an AI voice model that runs on your own device. Six English voices, adjustable speed, MP3 or WAV download, no account.

🌐 Español

🔒 Private by design: your text is processed locally in your browser and never uploaded to any server.

Written and maintained by Last updated

A voice model that lives in your tab

Almost every free text to speech site is a front end for a paid cloud API. That shapes everything about them: a character limit so the free tier stays free, an account so usage can be attributed, a watermark or a lower quality voice for anyone who has not paid, and your text sitting on somebody else’s server because that is where the synthesis happened.

This page inverts the arrangement. The voice model itself is fetched once, about 92 MB in the quantized build, and from then on it runs inside your browser on your own processor. There is nothing to meter, so there is no character limit and no account. There is no server involved in synthesis, so your text does not go anywhere.

The trade is honest and worth stating: you pay a one time download and each generation uses your own device rather than a data centre. On a modern machine that is a fair deal.

Producing an audio file

  1. Type or paste your text into the box.
  2. Choose a Voice from the six available.
  3. Adjust Speaking speed if the default pace is not right for your material.
  4. Choose MP3 or WAV as the download format.
  5. Generate, listen back, and download.

The very first generation includes the model download and will take noticeably longer than the ones after it.

The reason there are only English voices

The model behind this page can produce more than English. The tool does not offer more than English, and the reason is a component most people never think about.

Text to speech is two stages. The first converts written text into phonemes, the actual sound units of a language, resolving spelling into pronunciation. The second turns phonemes into a waveform. Only the second stage is the neural model. The first is a rule based system, and the rules bundled here cover American and British English.

Point an English voice at German text and the phoneme stage applies English rules to German spelling, producing something that sounds like an English speaker reading German aloud without knowing any. That is worse than useless because it sounds fluent. Offering only the languages the pronunciation layer genuinely supports is the honest configuration.

If your text is in another language, translate it first with Translate Text and then read the English version here.

Speed is generated, not applied afterwards

There is a meaningful difference between generating speech at a faster pace and speeding up a recording after the fact.

Playing audio faster compresses the waveform in time, which raises every frequency in it, which is why a sped up voice sounds high and thin. Some tools work around that with pitch correction, which introduces its own artefacts.

Here the speed setting is handed to the model before generation, so the voice speaks faster in the way a person does, with shorter pauses and quicker delivery, at its own natural pitch. The audio you download at 1.3 sounds like the same speaker in a hurry rather than a different, smaller speaker.

If you do need to change the speed of audio you already have, that is a separate operation and Change Audio Speed is built for it.

Two output formats and what to do with them

The model outputs raw single channel audio at 24 kHz. WAV is that audio wrapped in a container with nothing else done to it. MP3 adds one encoding pass at 192 kbps, which makes the file several times smaller at a quality difference you will not hear on speech.

Take WAV if the audio is heading into an editor, since re-encoding an already lossy file at the end of a chain is the thing to avoid. Take MP3 for anything you are simply going to play or send. If you take WAV and later want it smaller, WAV to MP3 does the conversion with a bitrate you choose, and Compress MP3 reduces an existing MP3 further.

For the reverse operation, turning recorded speech into text, Transcribe Audio runs a different model in the same local way.

Frequently asked questions

Where does the synthesis happen?

In your browser tab. The voice model is downloaded once from a public model host, roughly 92 MB in its quantized form plus a small style file per voice, and after that every conversion runs on your own device. Your text is never transmitted, there is no account, no daily quota and no watermark. Once the model is cached you can keep generating audio with no network connection at all.

Why are the voices English only when the model supports more?

Because the part that turns letters into sounds is the limiting factor, not the model. Producing speech requires converting written text into phonemes first, and the phoneme rules bundled here only cover American and British English. Feeding Spanish or German text to an English voice would produce confident sounding nonsense, so those voices are simply not offered rather than being offered and mispronouncing everything.

How were the six voices chosen?

They are a subset of the model's full built in voice set, restricted to those with the highest quality grades in the model's own published ratings, so none of the options sounds obviously worse than the others. Beyond that it is subjective. Try two or three with a paragraph of your own text, because voices differ noticeably over a long passage in ways a single sentence will not reveal.

What does the speed setting do to the audio?

It changes how the model speaks rather than how the finished audio is played back. That is an important distinction, because speeding up a recording afterwards raises the pitch and produces the familiar chipmunk effect, whereas generating at a higher speed keeps the voice sounding like itself. The range runs from half speed to double, with 1 as the model's natural pace, and a value outside that range is clamped back into it.

MP3 or WAV, and what is the difference here?

The model produces raw audio at 24 kHz in one channel. WAV wraps that directly with no further processing, so it is the exact output and the larger file. MP3 runs one additional encoding pass at 192 kbps, which is smaller and universally playable. Choose WAV if the audio is going into an editor for further work, and MP3 for anything that will simply be listened to or attached to something.

How good is it compared with a paid cloud voice?

Close enough to be useful and not indistinguishable. It is a large step beyond the flat system voices built into most operating systems, with real intonation and sentence rhythm. Where it still shows its size is on unusual names, abbreviations and numbers, which it will occasionally read in a way you did not intend. Listen back before publishing anything that matters.

Why is the first run slow and the rest fast?

The first run has to download the model, which is the 92 MB step, and initialise it. After that the model stays in the browser's cache and later conversions start immediately. Generation itself is a single pass through the model per utterance rather than a step by step loop, which is the main reason this runs at a reasonable speed in a browser at all.

Related tools