A pronunciation coach that won't lie to spare your feelings

What it does AI Pronunciation Trainer records your voice, transcribes speech with Whisper, and compares your phonemes against a reference to score how closely you pronounced a target phrase. It runs locally as a Python web app or through a hosted online version for those who would rather not manage ffmpeg and dependencies.

The interesting bit The tool is designed to solve a specific social failure mode of language learning: humans get bored and start saying “it sounds good” to be polite. It swaps that fragile feedback loop for Whisper transcription plus Epitran phoneme mapping, scoring your accent with algorithmic consistency. The architecture is modular enough that adding a new language is mostly a matter of dropping in a CSV database and wiring up a language code.

Key highlights

• Uses Whisper as the default ASR model; unsupported languages can be added by implementing the IASRModel interface in models.py.
• Phoneme-level scoring is handled by Epitran in RuleBasedModels.py, with per-language wrappers for character cleanup.
• Recent updates removed the OR-Tools dependency and relaxed Python requirements down to 3.12.
• Target text is editable before evaluation, so you aren’t locked to preset phrases.
• Adding a language requires touching callback.js, main.html, and a few backend files—a bit manual, but the author claims it is “doable with relative ease.”

Caveats

• The hosted version officially supports only desktop Chrome and imposes a daily usage cap.
• The author explicitly warns this is a spare-time project with small bugs still present.
• Language support is technically extensible, but you’ll need to edit both frontend and backend code by hand.

Verdict Try it if you want brutally honest pronunciation drills without burning social capital. Skip it if you need a polished, zero-config consumer app or mobile browser support.