Vocal Syllable Mapper

A research project on a question that sounds simple and is not: given a recording of someone singing or rapping, where exactly does each syllable land? Version 1 chased onset attacks, which is the wrong acoustic signal, and trained a model on its own detector's guesses. Version 2 is a rebuild around the vowel nucleus, the one part of a syllable that is always there, trained on 100 acapellas whose labels are anchored to a known syllable count derived from the lyrics. It now also finds word and line boundaries, which was the actual point of the project. The live demo lets you record or upload a track and scrub a waveform with a click track on every detected syllable.

Attention: Still in development, and the write-up is honest about where it is weak. Recall trails precision, no full song has true hand-labeled ground truth yet, and the line and section boundaries are the least reliable output on the page.