Suzune
Link to open source: https://github.com/magic-bubblez/suzune
you know when you wanna practice singing a song and you sing along with the original, in your head you sound so great because it feels good. but when you hear your own recording you get to know how bad you are. now there is no way of objectively seeing what your voice is doing.
that's where suzune comes. suzune is a vocal trainer that compares how well you are doing against the original song.
Upload a song. Demucs separates the vocal from the music; the vocal's pitch, loudness and lyrics are extracted and cached in a local library, so a song is processed once.
The original vocal is drawn as a graph: each thin white line is one moment of singing, its height is the note, the pink underneath is loudness. Zoom into any moment with the scroll wheel, drag to pan, click timestamps to jump.
Record live (with a live pitch monitor showing your note and how many cents you are from the original) or upload a pre-recorded vocal file.
Your take is aligned to the original with Dynamic Time Warping and scored.
Suzune, the coach, receives the pitch analysis and both transcripts (the original lyrics and what you actually sang) and writes a short report: verdict, the one thing to fix first with a drill, smaller notes, a real strength, and a focus for the next take. Timestamps in the report are clickable and play the original from that point.
Ask Suzune anything, at any time, in the chat at the bottom. The conversation is kept across takes.
Share a performance card as a PNG.
This build was uploaded as a hackathon project

