English Communication Trainer

The problem

Practising spoken English alone has no feedback loop. Apps either score pronunciation and nothing else, or send my audio to a service I do not control.

How it works

• Record in the browser. The backend transcribes with WhisperX on a local GPU and finds word timings, pauses and filler words.
• All the numbers are computed in code, not by a model: words per minute, pause map, filler counts, target-word matches, a weighted score. Re-running a session gives the same numbers.
• Claude only does the judgement part, working from the pre-computed summary instead of raw audio, which keeps the token cost low.
• Vocabulary to practise is scheduled with spaced repetition.

A decision: Hand work to Claude through a queue, not an API call

Pressing Process marks a session as pending and transcribes it. I then run one command in Claude Code, which works through every pending session. The app holds no API key and a web page never calls a model.

What was hard

Whisper-style models tidy up speech and drop filler words, so a transcript alone says the delivery is cleaner than it is. A separate voice-activity pass (Silero) measures silence from the audio and is compared with the transcript.