Pronunciation Coach — AI Lab

How it works

Pick a sentence. Most of them hide a classic learner trap: a long vowel, a small っ, or a word that turns into a different word if you get it slightly wrong.
Load the speech model. This downloads Whisper, OpenAI’s open speech-recognition model, plus a Japanese dictionary, into your browser. It’s about 95 MB the first time, then cached.
Record yourself. Whisper transcribes what you said, on your device. Your voice is never uploaded.
A dictionary-based analyser (kuromoji) turns Whisper’s transcript into kana. The page lines it up against the target sentence one kana at a time, highlighting what went missing and what appeared instead.

Don’t have a microphone handy? Press test with the reference voice to run the same pipeline on the model recording.

The catch: Whisper hears words, not sounds

I tested this with deliberately mispronounced recordings, and the results say a lot about how speech recognition works:

Say びよういん instead of びょういん, and the small model writes 美容院, beauty salon. That’s exactly the confusion a Japanese listener would have. The coach catches it.
Say まろ instead of まど, and it writes マロ. Caught.
Say ともたち instead of ともだち, or えが instead of えいが, and Whisper quietly writes 友達 and 映画 anyway. It was trained to produce likely Japanese, so small slips get autocorrected into the word you obviously meant.

So treat this as a “would a listener understand me?” check, not a phonetics lab. A perfect score means your words came through as the right words. It says nothing about pitch accent or how native you sound. When it does flag something, though, it’s usually a mistake that changes the meaning, and those are the mistakes worth fixing first.

Things I learned building it

Whisper imagines speech in noise. When I tested the recorder with a synthetic beep instead of a voice, Whisper confidently wrote さて、今の動画を見てくれてありがとう！: “well, thanks for watching this video!” Much of its training audio came with video subtitles, and those sign-offs are what it learned noise sounds like. The coach now checks the volume before transcribing, and catches these phantom sign-offs instead of scoring them.
Small vs. base is a real trade-off. The small model is almost three times the download but noticeably sharper on near-miss words. That’s why it’s offered as an option rather than the default.
The reference voice is AI too. The model recordings were generated with Gemini’s text-to-speech. Before shipping, I ran every one through Whisper to confirm each transcribes back to exactly the target kana.