Can a prompt replace ELSA Speak?

Build a local pronunciation practice loop for a language you are learning, and be precise about what it can and cannot tell you.

The loop: display a target sentence, record from the microphone, transcribe locally with whisper.cpp, and compare the transcript to the target with a word-level alignment. Show each word coloured by whether it was recognised, substituted or dropped, and keep the audio file beside the attempt so you can listen back.

Be honest in the interface about what this measures. A word the recogniser missed is evidence that something is off; a word it caught is not evidence that you said it well, because recognisers are trained to be forgiving of accent. Put that sentence in the UI, not in a README, because the temptation to read the output as a pronunciation score is strong and wrong.

Run each attempt through the recogniser three times with different temperature settings and report agreement across runs. A word that is transcribed differently on each pass is genuinely unclear in your recording — that instability is a better signal than any single transcript.

Build the error report over time: words dropped or substituted most often across all your attempts, with links to the recordings. That list is what you take to a tutor.

Do not compute a phoneme-level score. Forced alignment tools exist and will happily produce numbers, but calibrating them against native speech is the actual research problem here, and an uncalibrated score is a confident number about nothing.