How to turn text into speech and get a file you can keep

The difference between speaking and producing a file

Every modern browser has speechSynthesis. It reads text aloud using the voices installed on the machine, it costs nothing, and it works offline.

It also cannot give you a file. The API speaks through the audio output and exposes no buffer, no stream and no way to capture what it played. This is not an oversight: the voices belong to the operating system, and their licensing generally does not permit redistributing the audio.

So a read-aloud button is easy and an audio file is a different problem. If you want something you can put in a video, send to someone, or listen to on a walk, the browser's built-in speech will not get you there.

The three ways to get a file

A cloud API. Google, Amazon, Microsoft, ElevenLabs. Best quality, particularly for expressive delivery, and priced per character: typically a few dollars per million. Your text goes to their servers, which is worth thinking about before narrating anything confidential.

Recording the system voice. Audio Hijack, Soundflower, or a virtual audio device. Works, sounds like the built-in voices, and is a fiddly manual process per clip.

A model running locally. A neural voice small enough to run in the browser. The weights are a real download (tens of megabytes) cached afterwards, and then it is free, unlimited, offline, and the text never leaves your machine.

The third is what this tool does, and the honest trade is: not as good as the best paid voices, considerably better than the built-in ones, and free at any length.

What makes synthetic speech sound wrong

Modern neural voices are good enough that the failures are specific rather than general:

Homographs. Read, lead, live, bow, wind, tear. The pronunciation depends on meaning, and the model guesses from context. It usually gets it right and, when it does not, rewording is faster than fighting it.

Numbers and abbreviations. 1984 might be "nineteen eighty-four" or "one thousand nine hundred and eighty-four". St is Saint or Street. Dr is Doctor or Drive. Writing them out removes the ambiguity.

Acronyms. NASA is a word, FBI is letters, and nothing in the spelling says which.

Long sentences. Intonation drifts across a very long sentence because the model is planning prosody over a limited window. Shorter sentences sound better, which is also true of writing generally.

Missing punctuation. Commas and full stops are where pauses come from. Text without punctuation is read as an unbroken rush.

Writing for the ear

Small changes that help a lot:

• Spell out anything ambiguous: Doctor Smith, twenty twenty-six,

fifteen per cent.

• Break long sentences. If you would need a breath, so does the voice.
• Use full stops rather than semicolons, which get a shorter pause than the

sense usually wants.

• Read it aloud yourself first. Anything you stumble over, the model will too.
• Put a comma where you want a beat. It is a blunt instrument and it works.

Speed, and the file you get

Speaking rate around 130–150 words per minute is comfortable for listening. Faster is fine for material you already know, and tiring for anything new.

For the file itself: WAV is uncompressed, lossless, and large, around 2.8 MB per minute at 24 kHz. MP3 is roughly a tenth of that and lossy, which for speech is nearly always an acceptable trade.

Keep WAV if the audio is going into an edit and will be processed further. Use MP3 for anything that is just going to be listened to.

When a person is still the right answer

Synthetic speech is genuinely good for drafts, for accessibility, for proofreading by ear, for internal material, and for anything where the alternative is no audio at all.

For a flagship advert, an audiobook, or anything where the delivery is the product, a human reader is still better and it is not close. The gap is not in the sound of the voice any more; it is in the choices about emphasis that come from understanding the text.