What should a pronunciation app actually measure?

I’m Levin Dong, working on Parle at Tingnova Inc. While building it, I kept running into a product-language problem: “feedback” is often treated as one thing, even though apps may be showing very different signals.

Here is the distinction I now use:

1. Recording and replay preserves what the learner said so they can listen again.

2. Speech-recognition word detection reports which words a recognizer detected.

3. Acoustic or phoneme analysis measures defined properties of the audio signal.

4. Qualified human feedback adds contextual judgment from a trained reviewer.

5. Generative AI coaching can guide another attempt, but it can also be wrong.

These signals answer different questions. If a system recognizes the expected words, that can help the practice flow, but it does not by itself prove pronunciation quality, accent quality, fluency, intelligibility, or proficiency.

If a product displays a score, I think it should also say exactly what input was measured, what output is being reported, what evidence supports it, and what failure modes are known.

For Parle, we deliberately keep the boundaries visible: reviewed model audio and recording support listen-and-compare practice; Phrase Match reports which words in the model phrase speech recognition detected. It is not a pronunciation or accent score, and recognition can be wrong. AI Coach Léo is optional and can also be wrong.

I wrote a longer seven-check French pronunciation app selection guide with a ten-minute comparison exercise:

https://getparle.app/blog/french-pronunciation-feedback-app/

Question for other makers and learners: which feedback output earns your trust fastest—playback, detected words, a transparent acoustic measure, expert human notes, or guided AI prompts? What disclosure would you want beside it?