Toward Part-Aware Choral Transcription with singing voice assignment

Toward Part-Aware Choral Transcription with singing voice assignment. By Hanyu Meng and three others, from the University of New South Wales, the University of Western Australia, and two other institutions. Published on the seventh of October. Its code repository is worked out: about sixty-five source files, with a readme and a licence.

Automatic music transcription for a single instrument has advanced a great deal. A choir is a different listening problem. Choral work needs soprano, alto, tenor and bass, the four SATB parts, written down as separate lines, because that separation is what rehearsal, teaching and rebuilding a score depend on. Recent note-level choral transcription does not deliver it. It produces one merged track of notes, with every voice folded together, and that single track limits exactly the uses a choir has for it.

The authors introduce Part-aware Choral Transcription, called PawCT. To their knowledge it is the first end-to-end neural framework that does both jobs at once: it identifies which SATB parts are active in the choral audio, and it transcribes each of them into its own note-level track. Transcribing the notes and assigning them to a voice are treated as one problem, not two steps run in sequence.

Several pieces work together. There is part-specific prediction of onsets, offsets and frames, so each voice gets its own account of where notes begin, end and sustain. There is part-presence estimation, which decides whether a voice is singing at all. There is union-level supervision. And there are structured training targets built on a range prior, shortened to RP, which draws on the pitch ranges of the four SATB voices. An extension called ordered continuity, or OC, adds two further expectations: melodic continuity within a part, and an ordering of pitch across the parts.

The main test is on YouChorale. There the full version, PawCT-RP-OC, reaches a macro part-aware note F1 of zero point two two five at a fifty-millisecond onset tolerance. An adapted choral baseline scores zero point one six five, so the gain is thirty-six point four percent in relative terms. A two-stage pipeline that assigns parts after the fact reaches zero point one seven five.

There is also a part-agnostic variant, PagCT, which ignores voice labels. Its onset F1 at fifty milliseconds is zero point three eight two, against zero point two three seven for the previous state-of-the-art choral transcription model. Cross-dataset evaluations on CSD and Cantoria further assess how the approach holds up under dataset shift. Code and demos are released.

The comparison that carries the argument is the one against the two-stage pipeline. Doing transcription first and voice assignment afterwards scores below the joint model, and the authors read their results as showing the benefit of modelling the two together. For a choir, which voice sang a note is part of the note, and PawCT is built to recover both in a single pass.