protodemo voice

How small can a voice get?

Type a sentence and hear it read back — once by a model the size of a few photos, once by one nearly four times bigger. No expertise required to judge this one. Your ear does the whole job.

cooking · work in progress
Step one

Load a voice

Both download into this tab and speak on your own machine. The smaller one lives on this site.

Step two

Give it something to say

The sentences above are picked to be awkward on purpose — a question, a number, a name, a dash. Flat statements hide the difference; those four expose it.

Step three

Listen back

Load a voice above, then press Say it. Each take stays here so you can play them one after another.
This is a work in progress. Both voices are English only, and neither has any idea what it is saying — they turn spelling into sounds, with no notion of meaning. That is most obvious when a sentence needs emphasis to make sense.
What the two voices cost you
VoiceDownloadTo speak a sentenceWhere it lives
Kitten Nano23.8 MB~1.3 sthis site
Kokoro86.0 MB~4.5 sa public model hub

Those timings are from a browser, which is the number that matters to you. The same two models run about three to four times quicker outside one, because a web page gets a single processor core to work with unless it asks for more in a way that would break fetching the models in the first place. Nothing here is using a graphics chip yet; that is the obvious next improvement.

Both clear the bar for being understood. What the extra 62 megabytes buys is not clarity but ease — breath, rhythm, the sense that a sentence is heading somewhere. Whether that is worth four times the download is a judgement only you can make, which is the point of putting them side by side rather than quoting a score.

Neither model reads letters — and what that costs

A voice model does not work from spelling. It works from sounds, written as phonetic symbols, so something has to convert "forty five" into fˈɔːɹɾi fˈaɪv before either model sees it. That converter is a separate 1.3 MB piece of software doing it by rule and dictionary.

It is also the part most likely to be wrong. Give either voice an unusual surname, a street name or an abbreviation and the mistake usually happens here, before the model that gets blamed for it has done anything at all.

How I checked they are intelligible without trusting my own ears

Judging a voice by listening is subjective, and the person who built the page is the worst judge of whether it works. So both voices were fed a sentence and the audio handed to a speech recognition model, which had never seen the original text, to write down what it heard.

Both came back as "The appointment lasted 45 minutes" — the words intact, the numerals written the way a transcriber writes them. That does not prove either voice is pleasant. It does prove neither is producing mush, which is the part I could otherwise have fooled myself about.

Does anything leave my browser?

The model files download the first time you pick a voice, the way an image or font does. After that, the speaking happens on your machine.

What you type is never uploaded — there is nowhere for it to go, since this page has no server of its own. The audio is made in the tab and thrown away when you close it.