Natural text to speech on your own device: how it works and how clear it is

What makes a computer voice sound natural

A computer voice can say every word right and still sound flat. A natural voice has to do two jobs well: work out how each word is said, and then give the sentence a rhythm, with stress on the right words and a pause where a person would breathe.

The Text to Speech tool on this site does those jobs in two steps, both on your device.

1. Words to sounds. Your text is first turned into a list of speech sounds by a port of misaki, the pronunciation system the voice model was trained with. It looks words up in dictionaries of more than 180,000 English words and names, applies rules for numbers, dates, money, clock times and abbreviations, and sends words it has never seen to a small network that guesses from the spelling.
2. Sounds to speech. Those sounds go to Kokoro-82M, an open-source voice model with 82 million parameters, published under the Apache 2.0 licence. It produces the audio, with the rhythm and stress learned from recordings of people reading.

The first step decides more than it seems to. “Read” in “I read it yesterday” and “I will read it” is spelled the same and said differently, and “4:45 p.m.” or “£52” has to become words before any voice can say it. A wrong guess at this step gives a wrong word, however good the voice is.

How I checked that it is easy to understand

Saying a voice “sounds natural” is easy. I wanted a number that someone else could check, so the voice was measured the way speech systems usually are: by whether a listener writes down the right words.

The listener was Whisper small.en, an open speech-recognition model. It transcribed audio of 118 sentences that were never used to tune anything: 100 Harvard sentences, the test sentences the IEEE published in 1969 for measuring speech quality, and 18 news-like sentences with numbers, dates and money. The word error rate is the share of words the listener got wrong.

Voice	Word error rate	Sentences heard exactly right (of 118)
This site’s voice, as the page makes it	1.09%	110
The original model in Python, full precision	1.54%	109
macOS say, the system voice on a Mac	1.81%	102

The rule before the voice went on the site was simple: it ships only if its error rate is no worse than the system voice’s 1.81% and within 0.3 points of the original model. It passed. The copy of the model the site uses keeps its weights in 8 and 16 bits instead of 32, to keep the download down, and the test shows the smaller file is as clear as the original. The gap between 1.09% and 1.54% is about five words in 1,105, too few to call the smaller file better.

What the table cannot tell you is whether you like the voice. Each of the six can be heard before the download through its Sample button, so judge that part by ear.

Why it runs on your device, and what that costs

A server that makes speech has to be paid for by the minute. Here the voice runs in your browser, so the text you paste is never sent anywhere, and since your own device does the work there is no daily limit and no watermark on the audio. The cost is paid in two places.

A one-time download. The voice is about 105 MB for one accent, or 119 MB if no other AI tool on this site has run in your browser (the AI runtime is shared). It goes into your browser’s storage the first time and loads from there on later visits, so it is downloaded once.

Your device’s time. Every sentence is computed on your own processor. A fast recent computer needs slightly more than sixty seconds of work for sixty seconds of speech. Phones are slower, and I have not timed one on real hardware, so the page makes no claim about phones: it measures the speed of your own device as it works and shows the time left from that.

You do not have to wait for the whole recording. Playback starts with the first finished sentence while the rest are still being made, and a highlight moves through your text one sentence at a time, keeping pace with the voice. If you press Stop, the audio made so far stays ready to play or save, and Continue picks up at the sentence after it.

Choosing a voice and a speed

There are six voices: four American (Heart, Bella, Michael and Fenrir) and two British (Emma and George). The British Text to Speech page opens with Emma chosen. Emma and George look every word up in a separate UK English dictionary, so words such as “garage” and “vitamin” are said the British way.

The speed has four steps: Slower, Normal speed, Faster and Fastest. A slower setting helps for language practice or for a listener who finds fast speech hard; the faster settings suit a draft you are proofreading by ear.

Writing text that reads well

A few habits make a big difference to the result.

1. Write numbers as you want to hear them. Clock times, prices, dates and common abbreviations such as “Dr.” have their own reading rules. “Try a sample” on the page fills in a short text with a doctor’s title, a flight time, two prices and a seat count, so you can hear those rules at work. A product code or a unit such as km/h has no rule, so spell it out in full.
2. Spell unusual names as they sound. A name the dictionaries do not know goes to the spelling network, which sometimes guesses wrong. If it does, write the name as it is pronounced and make the audio again.
3. Use punctuation for pauses. Commas and full stops reach the voice model along with the speech sounds, and it pauses at them, so a list typed without commas runs together.
4. Split long texts. One run holds 20,000 characters, roughly 3,500 words. A long chapter therefore needs two or three runs.

Saving the audio

When the audio is made you can save it in two forms:

Format	What it is	Size and use
MP3	64 kbps compressed audio	About half a megabyte a minute; plays on every phone, player and editor
WAV	Uncompressed 16-bit audio	Much larger; best when you will edit the audio

The saved file takes its name from the opening words of your text, such as tts-welcome-to-the-tuesday.mp3 for the sample. On an iPhone or an Android phone, Save hands the file to the phone’s share sheet, where Save to Files keeps it and any app on the list can receive it. On a computer it opens Save As or downloads the file, depending on the browser.

Arabic, Urdu and other languages

The six AI voices speak only English: I have found no Arabic or Urdu voice model whose licence allows its use here. If your phone or computer has its own Arabic or Urdu voice, the page offers that voice instead, and reading starts straight away. That voice plays but cannot be saved as a file, because browsers do not give websites the audio of system voices. On an iPhone you can add voices in Settings, under Accessibility, Spoken Content, Voices.

If you need the opposite job, a recording turned into text, Audio to Text does that on your device too.