All projects

Project 07 of 12

Narratoqo

Audiobook narration that runs in the browser.

Status
Awaiting domain
Stack & integrations
  • React
  • TypeScript
  • Vite
  • Transformers.js
  • ONNX Runtime Web
  • Vercel
Fig. 1Reader1440 × 900 px
Narratoqo reader playing the Alice in Wonderland sample with the on-device Kokoro voice: the first segment is outlined, the spoken word is highlighted and the player shows a pause button
Web app build, captured November 6, 2026 · public-domain sample chapter, voice generated on the deviceView full capture

01What it does

A free web app that reads pasted text, articles and documents aloud with neural voices that run on the device, highlighting each word as it is spoken. No account or API key is needed, and projects and audio stay in the browser.

02The problem

Listening to an article or a chapter usually means a paid text-to-speech service or sending the text to someone else’s server.

03How it works

The open-source Kokoro and Piper voice models run in the browser through WebAssembly. The first play downloads the model once (about 93 MB for Kokoro), and the page shows that size before downloading. A one-click sample chapter plays without any setup. Cloud voices are optional and use the visitor’s own key. Narratoqo is awaiting its branded domain, and the launch link appears here once that domain is live.

04On-device inference

Kokoro speech is generated with ONNX Runtime in a worker, so the page stays responsive while a segment is synthesized. Text is split into segments that are generated on demand, cached in IndexedDB and played with word-level highlighting. In a headless test on a desktop CPU over a home connection, the first audio started about 9 seconds after the click on a first visit, including the model download, and about 4 seconds once the model was cached.

05Technical notes

  • Kokoro 82M (12 English voices) and Piper (18 voices in 10 languages) running locally through WebAssembly
  • Download size shown before the first play, with byte progress and browser caching
  • Import from pasted text, a web address, or TXT, Markdown and HTML files
  • Content Security Policy limited to version-pinned runtime paths and the Hugging Face model hosts, checked by tests