Pronunciation test
Pick a word, press record, say the phrase. Three speech-to-text models of different strength write what they heard. The goal: a model that is near-perfect on native speech but still trips on accent.
ship
Tap to start, tap again to stop.
You vs native speaker
Each cell: what that model heard from you (top) and from a native reference voice (bottom). The reference is recorded once and reused, so its numbers don't jump. ▶ plays the recording. A useful model has the native line all green.
| Word | whisper-tiny | whisper-base | large-v3-turbo |
|---|---|---|---|
All attempts
| Target | tiny | base | turbo | Audio |
|---|
Audio goes only to Cloudflare Workers AI and is not stored. "Hear it" plays the native recording.