Question 1

Why do some voices sound robotic and others do not?

Accepted Answer

Because they are technologies from different eras. Voices badged "Basic" are concatenative or formant synthesizers your operating system has shipped since the 2000s: they assemble speech by gluing recorded fragments, which is where the metallic sound comes from. Voices badged "Natural" are already neural models from the system itself (Microsoft's "Natural" voices, Google's on Android and Chrome). Those badged "Neural" and "Premium" are models this tool downloads and runs in your browser, and they are the best sounding. If the current voice is ugly, the fix is to filter by "Neural" and pick one in your language.

Question 2

Why do I need to download a large file to use the neural voices?

Accepted Answer

Because the model runs on your device, not on a server of ours. A Piper model is 20 MB to 110 MB, and Kokoro is around 86 MB. The download happens once per voice and the file is kept in browser storage (OPFS), so from the second time on the voice loads instantly, even offline. That is the price of never sending your text anywhere. You can see what is stored and delete what you no longer use in the "Advanced" panel.

Question 3

Why can I not download an MP3 with my system voice?

Accepted Answer

Because the browser never hands that audio to the page. The Web Speech API exposes playback commands only: the page asks the system to speak and it speaks, without ever passing the audio samples back. That is not a limitation of this tool, it is how the standard was designed, and it applies to any website. To get a file, pick a Neural or Premium voice: those run inside the page, so the audio exists here and can be exported as MP3 or WAV.

Question 4

How do I make the voice pause, emphasise or slow down in one passage?

Accepted Answer

Use the markup bar above the text, or type it directly. Two notations are accepted: brackets, like [pause 800ms], [slow]text[/slow] and [whisper]text[/whisper]; and SSML-style tags, like , text and text. No browser actually implements SSML, so the tool itself interprets the markup, which is why it works identically across all three engines. An unknown tag becomes a warning and is dropped, never read aloud.

Question 5

Does the tool read numbers, dates and money correctly?

Accepted Answer

Yes, and this is one of the parts that changes the result the most. Before synthesis the text goes through normalization that spells out what is spoken out loud: "$1,234.56" becomes "one thousand two hundred thirty-four dollars and fifty-six cents", "03/12/2026" becomes the full date, "120 km/h" becomes "one hundred twenty kilometers per hour", "3rd" becomes "third" and "21st century" is read as words. Each rule can be turned off individually in the "Advanced" panel if your text needs to be read literally.

Question 6

Can I build a dialogue with a different voice per character?

Accepted Answer

Yes. Write the text as "Name: line", one line per turn, and the tool detects the characters on its own. A voice selector for each one appears in the "Advanced" panel. The character name is not spoken: the voice change is what marks the turn. The detector is deliberately conservative, so labels like "Note:" or "Warning:" inside ordinary text do not become characters.

Question 7

Do the SRT subtitles actually line up with the audio?

Accepted Answer

They do, because the timings are measured rather than estimated. Since synthesis happens here, we know the exact duration of every spoken chunk in the generated file, and the subtitles use those numbers. That is unlike tools that estimate from "words per minute", which accumulate error and drift by seconds after a few minutes of audio. Inside a very long sentence split across more than one block, the time is divided in proportion to character count, but the start and end of each sentence are exact.

Question 8

What is the character limit?

Accepted Answer

There is no fixed limit. The text is split into sentences and each sentence is spoken as a unit, which fixes the classic Chrome stall on long text sent in a single call. With the system voice, a large text only costs time. With a neural voice, each sentence is synthesized while the previous one plays, so audio starts within seconds even on a multi-page text, and export only waits for the whole text to finish generating.

Question 9

Is my text sent to any server?

Accepted Answer

No. The text, the generated audio, the pronunciation dictionary and your preferences all stay on your device. The only network requests the tool makes are to download the neural voice model files, hosted on public Hugging Face repositories, and those requests carry only the file request, never your text. If you use the system voice alone, no request is made at all.

Question 10

What does the pronunciation dictionary do?

Accepted Answer

It swaps terms before the text reaches the synthesizer, which fixes proper nouns, acronyms and loanwords the voice gets wrong. Writing that "SQL" should be read as "sequel", or respelling a place name phonetically, corrects the reading in any engine. Matching is whole-word, so "SQL" does not alter "SQLite", and longer terms are applied before shorter ones, so "SQL Server" is not fragmented by the "SQL" rule.

Question 11

Does it work on mobile and offline?

Accepted Answer

The system voice works on any modern phone and needs no internet. Neural voices need internet only the first time, to download the model; after that they work offline, as long as the browser has not cleared storage. On older phones a large model may not fit in memory: if that happens the tool says so, and you can pick a "low" or "x_low" quality voice, which are much smaller, or fall back to the system voice.

Question 12

Can I use the generated audio commercially?

Accepted Answer

On this tool's side, yes: we charge nothing and claim no rights over what you generate. What you need to check is the licence of the specific voice. Piper models come from varied datasets, each voice carrying its own licence in the rhasspy/piper-voices repository on Hugging Face, and some are restricted to non-commercial use. Kokoro is distributed under Apache 2.0. System voices follow your operating system's terms. Check the licence of the voice you picked before publishing.

Text to Speech Converter

Text to Speech

Three speech engines, all running inside your browser

Paste the text, pick a voice and adjust before hitting play.

Use this tool on your site

Sources and references for this tool

Frequently asked questions

Related tools

Text to Speech Converter

Voice

Controls

Reading

Export

Text to Speech

Three speech engines, all running inside your browser

Paste the text, pick a voice and adjust before hitting play.

Use this tool on your site

Sources and references for this tool

Frequently asked questions

Related tools

Voice

Controls

Reading

Export

Text to Speech

Three speech engines, all running inside your browser

Paste the text, pick a voice and adjust before hitting play.

Use this tool on your site

Sources and references for this tool

Frequently asked questions

Where to go next

Related tools

Voice

Controls

Reading

Export

Text to Speech

Three speech engines, all running inside your browser

Paste the text, pick a voice and adjust before hitting play.

Use this tool on your site

Sources and references for this tool

Frequently asked questions

Where to go next

Related tools

Voice

Controls

Reading

Export