Text to Speech Reader

Reading aloud as a proofreading tool

This is the most useful thing on the page, and it has nothing to do with accessibility features or convenience.

When you read your own writing silently, you do not read it. You reconstruct it. Your brain already knows what the sentence was supposed to say, so it supplies the missing word, smooths the clumsy clause and skips the repetition. This is why you can proofread a paragraph four times and still miss that "the the" in the middle of it.

Hearing the text removes the shortcut. The words arrive in sequence, at a fixed speed, with no opportunity to skip ahead, and several classes of problem become impossible to miss:

• Repetition. The same word three lines apart is invisible on the page and glaring in the ear.
• Missing small words. A dropped "of" or "the" is exactly what silent reading repairs automatically.
• Sentences that run too long. If a synthetic voice sounds like it is running out of breath, a reader is too.
• Clumsy rhythm. Three sentences of the same length and structure in a row have a monotony you can hear and cannot see.
• Accidental rhyme and alliteration. Fine in poetry, distracting in a report.

A synthetic voice is arguably better for this than reading aloud yourself, because it will not unconsciously add the emphasis that makes a badly built sentence sound intentional. It reads what you wrote, not what you meant.

Where the voices come from

Browsers expose a speech synthesis interface that hands the work to whatever voices your operating system provides. That is why the list differs between a Mac, a Windows machine and a phone, and why it can change after a system update.

Some systems distinguish local voices from network ones. Local voices work offline and start instantly; network voices usually sound considerably better but need a connection and may pause briefly before starting. If a voice sounds noticeably more natural than the rest, it is probably a network voice.

Because everything happens on your device, there is no length limit, no cost and nothing to sign up for. The trade is that the audio cannot be captured to a file: it goes straight to your speakers, and there is no output stream for the page to record. Any tool offering a downloadable MP3 is processing your text on a server.

The fifteen second bug

Several browsers, most persistently Chrome on desktop, stop speech after roughly fifteen seconds if a single utterance runs longer than that. It has been a known issue for years.

The workaround, which this page uses, is to split the text into sentences and queue them as separate utterances. Each one is short enough to complete, and the queue advances automatically.

This has two useful side effects. Pause and resume become far more reliable, since the engine only ever holds a short piece of text. And it becomes possible to highlight which sentence is currently being read, which turns listening into something you can follow along with rather than just hear.

Other things it is genuinely good for

Getting through long documents. Listening at 1.5x while doing something else covers far more ground than reading would. Comprehension holds up reasonably well for familiar material, less so for dense or technical text.

Accessibility. For people with dyslexia, visual impairment or fatigue, hearing text alongside seeing it measurably improves both reading speed and retention. This is not a substitute for a proper screen reader, which understands page structure, but it is useful for a block of text.

Language learning. Hearing correct pronunciation of written text, repeatedly and on demand, at a speed you control. Set the language filter and pick a native voice.

Checking dialogue. Written speech that reads fine on the page frequently sounds nothing like a person talking. Hearing it is the quickest test.

Getting a better result

Punctuate properly. The engine derives its pauses entirely from punctuation. A paragraph with no commas is read as one long breathless run, which is informative about your writing but unpleasant to listen to.

Slow down for proofreading, speed up for consumption. Around 0.9x makes errors easier to catch. Between 1.3x and 1.7x is comfortable for getting through material you broadly understand.

Expect names and acronyms to be mangled. Synthesis engines guess from spelling. If a name matters, spelling it phonetically in a scratch copy will get you a better reading.

Try more than one voice. Quality varies enormously within the same machine, and the default is rarely the best available.