Can a prompt replace Pimsleur?

Build an audio-first recall drill you can run without looking at a screen.

A drill item is a prompt audio file, a response audio file and a pause length. A session plays prompt, waits the pause, plays the response, and moves on. There is no interface during a session — only audio — because the whole point is that it works while walking or driving.

Scheduling follows the graduated-interval idea rather than an app-style queue: an item first heard today comes back a few seconds later, then a minute, then five, then in a later session, then several sessions on. Implement it as an explicit list of intervals in a config file so you can tune it, and store per-item position in that list.

Since there is no screen, grading is implicit: sessions are generated ahead of time as a single audio file per session, built with ffmpeg by concatenating clips and silences. That way a session is a file you can put on a phone, a watch or a car stereo with no app at all.

Recording is a small side tool: type a phrase pair, record both sides, save them with the pair. Recording your own target-language audio is only useful if you already speak it well; otherwise use audio you have licence to use, and say in the README where yours came from.

Do not use text-to-speech for the target language unless you have listened to it critically in that language. Synthetic voices are convincing in English and frequently wrong elsewhere, and a drill that rehearses wrong prosody several hundred times is worse than no drill.