ACappellaSet: A Multilingual A Cappella Dataset for Source Separation and AI-assisted Rehearsal Tools

Ting-Yu Pan, Kexin Phyllis Ju, Hao-Wen Dong

Keywords: A Cappella, Music Source Separation, Dataset, Data Augmentation

Abstract: A cappella music presents unique challenges for source separation due to its diverse vocal styles and the coexistence of harmonic and percussive voices. Current a cappella datasets are limited in size and diversity, hindering the development of robust source separation models. In this paper, we present ACappellaSet, a collection of 55 professionally recorded a cappella songs performed by three professional groups. In addition, we present experimental results showing that fine-tuning HTDemucs on ACappellaSet substantially improves vocal percussion (VP) separation, raising VP SDR from 5.22~dB to 7.62~dB, and enabling scalable multi-stem modeling. Finally, we discuss future work on AI-driven dataset augmentation and supporting tools for asynchronous a cappella rehearsals.

Track: Paper Track

Confirmation: Paper Track: I confirm that I have followed the formatting guideline and anonymized my submission.

Submission Number: 107