use case · multi-model · Whisper + Kokoro
Two browser models wired into one pipeline. Whisper turns your voice into text; you tweak the words; then Kokoro speaks them back in a voice you choose. Speech in, speech out — and nothing ever leaves the device.
This loop chains two on-device models — each initialises on its own below (cached copies auto-start, no button).
Whisper — speech-to-text
Kokoro — text-to-speech
Preparing Whisper and Kokoro on your device…
step 1 — Whisper
Bundled JFK inaugural excerpt (20 January 1961), spoken by John F. Kennedy — public domain, U.S. government work · JFK audio source and attribution record.
In headless / embedded views there's no mic — use the sample or upload a file. The loop is identical.
step 2 — you
step 3 — Kokoro
On their own, each model is useful. Chained, they're a template for a whole class of app: a private voice assistant, a pronunciation coach, a "re-voice this in a calmer tone" tool, a hands-free note-taker that reads your notes back. Drop an on-device LLM between steps 2 and 3 and you have a full voice agent — with no server, no API bill, and no audio ever leaving the tab.