Read this aloud. See what a rater would mark.

Sentence 1 of 6

Press to get the model ready: about 130 MB, once.

Downloaded once, then kept by your browser. Nothing you record is uploaded.

Your voice stays on this device. Nothing is uploaded, and your browser’s network tab will show you so.

How 1.2 GB became 123 MB

The Python library listens through a large speech model: about 315 million parameters, 1.2 GB on disk. A server carries that without trouble. A web page cannot.

So a smaller model, with 95 million parameters, was trained to imitate it. Both listened to the same 32 hours of English speech, with no transcript and no rating, and the small one learned to answer as the large one did. Rounded to 8-bit integers, its weights come to 123 MB, which your browser downloads once and runs with onnxruntime-web.

Shrinking it cost nothing measurable. On the 2,500 test recordings of speechocean762, each scored by expert raters, the small model finds the mispronounced words as well as the large one, and its total follows the raters a little more closely.

ModelSizeWord detection (AUC)Total score (Pearson)
Large, the library’s default1.2 GB0.9280.790
Small382 MB0.9380.802
Small, 8-bit, in this page123 MB0.9390.798

Read those numbers with one caution. While it learned, the small model heard speakers from the training half of the same corpus, so on your own recordings expect what the large model gives, not better.

The model on Hugging FaceThe JavaScript codeAll the benchmarks

In your own app, with no audio server

What this page does, your web app can do. The recording is marked on the speaker’s phone or laptop, so you have no audio server to run and nothing to pay per minute of speech. The analysis sends no recording to a third party, so it adds no processor to declare under the GDPR. Once the model is loaded, it keeps working offline.

Today it covers English read aloud. The JavaScript is under the AGPL, like the rest of OpenPronounce, and shipping it inside a closed product takes the commercial licence.

The page handles English read aloud only; free speech and the other languages are in the Python library. Its scores can differ from the library’s by a couple of points, because it runs the 8-bit weights.