One voice,
two ways to read.

Type Nepali and hear it read twice — once as a news reader, once as an audiobook reader.

research This is a research voice, not a broadcast reader. 3.2 million parameters, 16 kHz, and the current release candidate.

News

Brisk, front-of-desk. Higher pitched and quicker through a sentence — the way a bulletin is read.

Audiobook

Slower and long-form. Lower pitched, more evenly paced — the way you would want a chapter read to you.

what you are listening to

Both clips are the same voice and the same checkpoint. The only thing that changes is the reading style, which the model takes as an input: the text frontend, the pitch and duration prediction, and the random seed are all held fixed, so the two clips are the same sentence read two ways rather than two models. On our test sentences the audiobook reading comes out about 3–8% longer, roughly 60 Hz lower in mean pitch, and about a third quieter than the news reading.

The decoder is 3.2 million parameters and produces 16 kHz audio. Devanagari is read directly by our own Nepali text frontend rather than transliterated through English. Pronunciation of unusual proper nouns and of English loanwords is the weakest part, and known to be.

where the voice actually runs

In the Chrome extension, the voice runs entirely in your browser — the model is downloaded once and every sentence is synthesised on your own machine. No audio, no text and no page you are reading is sent anywhere. There is no cloud in the loop and no account to make.

This page is the exception, and it is worth being exact about it: this demo renders on our server, because it is a web page and it has to. What you type here is posted to a small machine in a data centre, read aloud there, and the audio is sent back. Nothing you type is stored. If you want the local version, that is the extension, not this box.

Up to 200 characters per go — the server renders at about real time on one shared CPU, so a long paragraph is mostly a long wait.

find us