Request access to ACappellaSet
Access to ACappellaSet is restricted to approved users. Please review and agree to the Non-commercial Research Use Terms of Use. All access requests are manually reviewed.
Before submitting your request, please read the full ACappellaSet Non-commercial Research Use Terms of Use (Version 1.0). Full Terms: https://huggingface.co/datasets/AcappellaSet/ACappellaSet/blob/main/LICENSE (also reproduced in full in the public dataset card below). By selecting the agreement checkbox, you acknowledge that you have read the Terms and agree to be legally bound by them if access is granted. Terms version 1.0 is effective October 21, 2026.
Log in or Sign Up to review the conditions and access this dataset content.
ACappellaSet v1.0
ACappellaSet is a multilingual multitrack a cappella dataset comprising 53 songs from three vocal groups. It provides isolated vocal-part stems and recording-level annotations to support research in music and audio.
The v1.0 package comprises two audio archives (primary recordings and alternates), with documentation, metadata, manifests, and checksums available as standalone files. Extract both audio archives into the dataset root, preserving their internal paths.
License and Access
ACappellaSet is available to individually approved users under the custom Non-commercial Research Use Terms of Use, regardless of institutional affiliation. Eligibility depends on the purpose of use. Access requires explicit acceptance of the Terms and Manual Approval; no separately signed DUA is required.
Read the complete Full Terms of Use in this public card before requesting access. The same text is provided in LICENSE; the separate file route may require approval. Redistribution of restricted audio is prohibited, and commercial use requires separate authorization. Research checkpoints and short research audio demonstrations are subject to Section 5. These Terms do not assign licenses to separately released code or model-weight files.
The contributing groups have authorized inclusion and distribution for Non-commercial Research Use within the rights they control. The Terms apply only to rights the dataset providers are authorized to license and grant no additional rights to third-party copyrighted works. The ACappellaSet Dataset Team administers access and does not claim ownership of third-party musical compositions or lyrics.
- Sign in to your own Hugging Face account and open AcappellaSet/ACappellaSet.
- Read the Terms, complete the request form, select both agreement checkboxes, and click Request access. Independent researchers may enter “independent researcher” for affiliation.
- Install the Hugging Face CLI if needed (
python -m pip install --upgrade huggingface_hub). After manual approval, authenticate with your approved account and download:
hf auth login
hf download AcappellaSet/ACappellaSet --repo-type dataset --local-dir ./AcappellaSet
Each collaborator needs individual approval unless the administrator authorizes another arrangement. Keep access tokens out of scripts and repositories.
Research access, questions, and corrections: pting@umich.edu, hwdong@umich.edu, or kexinju@umich.edu. For Seagull-K commercial licensing, contact seagullkvocalband@gmail.com; for other groups, contact their authorized representatives. The dataset team does not independently grant commercial licenses.
Email and affiliation: Please verify the email address associated with your Hugging Face account before requesting access. An institutional email address matching your stated university, research institute, or organization is preferred when available.
If you use a personal email address, please state your affiliation—or indicate that you are an independent researcher—and describe your intended research use. A public research profile or project webpage is helpful but optional. All access requests are reviewed individually.
Research Uses and Considerations
ACappellaSet supports a range of non-commercial music and audio research applications, including but not limited to source separation, music transcription, vocal analysis, and music information retrieval. All uses are subject to the Non-commercial Research Use Terms.
Researchers should consider the following when using the dataset:
- Singer generalization: The provided splits are not singer-disjoint and should not be used to claim generalization to entirely unseen singers.
- Annotation granularity: Singer identities and double tracking are not annotated with precise temporal boundaries.
- Related recording versions: Different versions of the same composition share musical material and must remain in the same experimental split.
- Dataset coverage: The recordings represent a limited number of vocal groups, languages, and recording conditions.
For further details, see METADATA_SCHEMA.md, PROVENANCE.md, and SPLITS.md.
Contents
| Component | Song families | Recording versions |
|---|---|---|
| Seagull-K (SK) | 47 | 100 |
| Annappella (AP) | 3 | 4 |
| REAL A Cappella (RA) | 3 | 3 |
| Total | 53 | 107 |
Seagull-K is a professional group; Annappella and REAL A Cappella are student groups. In the accompanying research, AP and RA are combined into the AB corpus (6 song families; 7 recording versions) because of their small sizes and overlapping membership. Group identifies the actual performing group; AB is an experimental corpus label.
There are 53 primary versions and 54 alternates: 44 produced, 39 edited, and 24 rough recordings across both sets. Audio comprises 647 WAV files (20,850,242,596 bytes), with 44.1-kHz two-channel recordings. All 540 role stems are 16-bit PCM. The 107 mixtures comprise 60 16-bit PCM files and 47 32-bit floating-point files. Sample encoding is specified per file in file_manifest.csv; the distributed stems are not 24-bit PCM.
Total duration across all recording versions is 5.0050 hours. Primary-version duration is 2.4989 hours. Alternate versions share musical material; total version duration is not unique musical duration. Two channels may contain dual-mono audio rather than a spatial stereo mix.
Roles are soprano, alto, tenor, baritone, bass, and vocal percussion (vp). Absent roles have no WAV file and are marked 0 in the curator's metadata/metadata.csv. Some role stems combine singers or layered takes and are not isolated individual voices.
metadata/metadata.csv is the curator-maintained main table with 20 fields; see METADATA_SCHEMA.md for complete definitions. All versions of each song share singer annotations, language, and original song name. The six singer fields identify performers associated with each vocal-part stem and may list multiple or alternating singers; annotations are not time-resolved, and 0 means the part is absent. composition_type distinguishes original compositions, covers, and medleys; an original arrangement of an existing song is still a cover. original song name preserves the title exactly as written in the arrangement score, which may differ from the underlying composition's official title. Reference splits remain in metadata/splits.csv.
Recording variant (variant)
Some songs are available in multiple recording variants, described by the variant column:
- Rough (
rough): Minimally processed recordings that largely preserve the original performance characteristics. - Edited (
edited): Recordings with corrective editing or processing, such as pitch correction, timing adjustments, or equalization. - Produced (
produced): Recordings with more extensive production processing, potentially including refined vocal editing, mixing, panning, reverberation, and other effects.
These categories describe the level and characteristics of audio production rather than a fixed sequential processing pipeline. Not all songs are available in all three versions.
Primary selection
For each song, the primary version is selected in this order of preference: produced → edited → rough. Use the produced version when available; otherwise use the edited version, or the rough version if neither produced nor edited is available.
Exactly one version per song is primary. The selected version has is_primary=1 in metadata/metadata.csv and is stored under audio/primary/<song_id>/. All remaining versions have is_primary=0 and are stored under audio/alternates/<song_id>/<variant>/.
Primary designation does not exclude alternate versions from training or evaluation. All versions of a composition must remain in the same train/validation/test split.
The current release has 53 primary recordings: 44 produced and 9 rough. No song currently requires the edited-version fallback. The remaining 54 recordings are alternates: 39 edited and 15 rough.
Recording Setup
The dataset includes three recording configurations, indicated by the recording_setup field in metadata/metadata.csv:
- Individual (
individual): Vocal parts were recorded separately, without simultaneous ensemble performance. - Simultaneous Local (
simultaneous_local): Singers performed simultaneously at the same physical location, potentially in separate rooms or partially acoustically isolated spaces. Some microphone leakage may be present. - Simultaneous Remote (
simultaneous_remote): Singers performed simultaneously from geographically separated locations using Jamulus for real-time communication and its built-in multitrack recording functionality. Recordings were captured through Jamulus rather than independent local recordings and may contain audio compression artifacts and network-related distortions, including packet-loss artifacts or audio dropouts.
Recording setup is specific to each recording version. Different versions of the same song may have different setups; the annotation describes each recording version.
spatial_type describes the resulting audio, primarily from post-production panning, rather than the recording setup: dual_mono has identical or nearly identical left/right signals with little or no stereo separation; stereo_spatial has meaningful spatial differences, typically from panning or other stereo processing. num_channels=2 does not necessarily mean stereo_spatial.
Double Tracking
Double Tracking (double_tracked): A recording-level binary indicator: 0 means no vocal double tracking is indicated; 1 means at least one vocal stem contains layered takes of the same singer performing the same part.
Double tracking is often used on accompanying parts, but the melodic lead may shift between parts across song sections, so its use is arrangement-dependent. The table does not specify which stems or time segments are double-tracked. Do not assume every stem in a flagged recording is double-tracked. Double tracking of one singer is distinct from a stem containing multiple different singers.
Language annotations
language describes the lyrics' language; most multi-language labels use +. PC01 retains the curator's exact comma-separated value MinNan, mandarin. CK03 (Welcome Horizons) has no lyrics and uses wordless in both versions. This label describes a vocal recording without lyrics, not an instrumental recording or a missing annotation.
Layout
The repository stores the two audio archives under audio/. After extraction into the dataset root, the working layout is:
audio/primary/<song_id>/*.wav
audio/alternates/<song_id>/<variant>/*.wav
metadata/metadata.csv
metadata/splits.csv
file_manifest.csv
checksums/audio_SHA256SUMS
See Research Uses and Considerations, METADATA_SCHEMA.md for field definitions, PROVENANCE.md for processing and training compatibility, and SPLITS.md before creating experimental splits.
Getting started
This release is distributed as downloadable audio archives and standalone metadata. The automatic Dataset Viewer is disabled; use the hf download command above. Direct automatic loading with datasets.load_dataset("AcappellaSet/ACappellaSet") is not configured.
| Archive | Contents |
|---|---|
audio/AcappellaSet_v1.0_primary.zip |
Audio for 53 primary recording versions under audio/primary/ |
audio/AcappellaSet_v1.0_alternates.zip |
Audio for 54 alternate recording versions under audio/alternates/ |
- Extract
audio/AcappellaSet_v1.0_primary.zipandaudio/AcappellaSet_v1.0_alternates.zipinto the dataset root, preserving their internal paths. The audio ZIPs already contain theaudio/prefix; do not extract them intoaudio/, which would createaudio/audio/. - Open
metadata/metadata.csvto find a recording by song ID and version. Primary recordings are inaudio/primary/<song_id>/; alternates are inaudio/alternates/<song_id>/<variant>/. - Each recording directory contains
mixture.wavand its available role stems. A singer field of0means that role is absent and no corresponding WAV is supplied. - Optionally verify downloaded files against
SHA256SUMSand extracted WAVs againstchecksums/audio_SHA256SUMS(see Integrity).
The hf download command above retrieves both audio archives and all standalone documentation, metadata, manifests, and checksums. No separate documentation archive is needed.
Primary and alternate versions share musical material. Keep all versions of a song together when forming train/validation/test splits; metadata/splits.csv provides one song_id,split assignment per song. Join on song_id to apply it to all recording versions; see SPLITS.md for the reference protocol and PC01 benchmark exclusion.
Some mixtures use floating-point WAV and contain sample values above ±1. Preserve floating-point values when loading or processing these files. Converting directly to integer PCM can clip them. If applying attenuation or normalization, apply the same gain to the mixture and all its stems. See PROVENANCE.md for mixture construction and signal-range details.
Integrity
Checksums are SHA-256 fingerprints that let you detect incomplete downloads or changed files. SHA256SUMS covers the downloadable archives and standalone release files; checksums/audio_SHA256SUMS covers the extracted WAVs; checksums/metadata_SHA256SUMS covers the standalone documentation and metadata.
Verification is optional and needs no project-specific scripts. From the dataset root on macOS, run shasum -a 256 -c SHA256SUMS to verify downloaded release files, or shasum -a 256 -c checksums/audio_SHA256SUMS after extracting the audio. On Linux, use sha256sum -c with the same checksum filename. A check requires the listed files to be present at their documented paths.
Each file is identified by SHA-256 and byte count. Full-waveform checks cover 106 mixtures; PC01 is outside this diagnostic coverage, separately from its benchmark exclusion; the largest absolute mixture-minus-stem-sum residual is 0. This is a dataset integrity check, not a model-performance evaluation.
Report corrections using recording IDs and filenames so changes can retain version and checksum provenance.
Citation and license
If you use ACappellaSet in your research, please cite our initial dataset paper:
Ting-Yu Pan, Kexin Phyllis Ju, and Hao-Wen Dong. “ACappellaSet: A Multilingual A Cappella Dataset for Source Separation and AI-assisted Rehearsal Tools.” NeurIPS 2025 Workshop on AI for Music (AI4Music), 2025. Non-archival workshop paper.
@inproceedings{pan2025acappellaset,
title = {{ACappellaSet}: A Multilingual A Cappella Dataset
for Source Separation and {AI}-assisted Rehearsal Tools},
author = {Pan, Ting-Yu and Ju, Kexin Phyllis and Dong, Hao-Wen},
booktitle = {NeurIPS 2025 Workshop on AI for Music},
year = {2025},
note = {Non-archival workshop paper},
url = {https://openreview.net/forum?id=rAPN9Ir7jt}
}
Please also identify the dataset version used, ACappellaSet v1.0, and link to the dataset repository.
The workshop paper describes an earlier dataset version. For the contents, statistics, audio formats, and annotations of the v1.0 release, refer to the documentation in this repository.
Dataset Credits: Ting-Yu Pan and Kexin Phyllis Ju. Acknowledgments to Seagull-K, Annappella, REAL A Cappella, and Hao-Wen Dong.
Restricted dataset materials are covered by the custom Non-commercial Research Use Terms below, version 1.0, effective October 21, 2026, subject to the stated third-party rights limitation.
Full Terms of Use
The following reproduces the complete LICENSE, version 1.0, effective October 21, 2026, for review before requesting access.
ACappellaSet Non-commercial Research Use Terms of Use
Version 1.0
Effective Date: October 21, 2026
Dataset Administrator:
ACappellaSet Dataset Team
Contributing Groups:
Seagull-K, Annappella, and REAL A Cappella
Commercial Licensing: See Section 3.
The ACappellaSet Dataset Team administers research access on behalf of the contributing groups, within the scope of their respective authorizations.
The contributing groups have authorized the inclusion and distribution of their recordings for Non-commercial Research Use under these Terms. This authorization applies only to rights controlled by the contributing groups and does not establish ownership or clearance of underlying third-party musical compositions or lyrics.
1. Scope and Acceptance
These Terms govern access to and use of the ACappellaSet dataset, including audio recordings, isolated vocal stems, mixtures, and any restricted accompanying data distributed as part of the approved release.
By explicitly accepting these Terms and receiving access approval, users agree to comply with the conditions set forth below.
Access is granted to approved individual accounts only. Institutional authorization must be established separately if an institution is intended to be bound by these Terms.
Subject to these Terms and applicable third-party rights, approved users receive limited, non-exclusive, non-transferable permission to access and use the dataset for the purposes specified in Section 2.
2. Permitted Uses
The dataset may be used solely for approved Non-commercial Research Use, subject to applicable third-party rights.
Non-commercial Research Use refers to research and educational activities conducted for non-commercial purposes.
Such use is permitted regardless of institutional affiliation, including universities, research institutes, for-profit companies, nonprofit organizations, and independent researchers.
Non-commercial Research Use does not include activities primarily intended to develop, improve, validate, or commercialize products or services for commercial advantage or monetary compensation.
Eligibility is determined by the actual purpose of use rather than the applicant’s institutional affiliation. All applicants are subject to the same manual access approval process. Company affiliation alone does not disqualify an applicant from Non-commercial Research Use.
Permitted activities include:
* Music source separation research.
* Machine learning training, validation, and evaluation.
* Audio signal processing and music information retrieval research.
* Reproducible research benchmarking.
* Internal preprocessing, augmentation, and analysis.
* Preparation of research publications and related results.
Internal copies and processed versions may be created only as reasonably necessary for permitted research.
3. Commercial Use
Commercial use is not authorized under these Terms.
Any proposed commercial use requires separate written authorization from the appropriate rights holders.
For commercial licensing inquiries involving Seagull-K recordings, please contact:
Seagull-K
seagullkvocalband@gmail.com
For recordings contributed by other groups, please contact their respective authorized representatives.
The ACappellaSet Dataset Team administers research access but does not independently grant commercial licenses.
Dataset access approval does not constitute commercial licensing approval.
4. Redistribution and Access Restrictions
Users must not:
* Redistribute, sell, sublicense, or publicly upload original recordings, stems, or mixtures.
* Publish processed copies that substantially reproduce or replace the original dataset recordings.
* Share restricted data or access credentials with unauthorized individuals.
* Make restricted audio publicly accessible through repositories, cloud storage, websites, or APIs, except for short research demonstrations permitted under Section 5.
* Use the recordings to develop or distribute systems intended to impersonate identifiable singers without separate authorization.
* Use the dataset beyond the permitted purposes.
Research collaborators must obtain individual access approval unless the dataset administrator explicitly authorizes another arrangement.
These restrictions do not prohibit the publication of research outputs expressly permitted under Section 5.
5. Research Publications and Model Outputs
Users may publish research papers, evaluation metrics, analyses, figures, and software implementations resulting from permitted research, provided that these outputs do not disclose restricted dataset recordings.
Publication and redistribution of trained model checkpoints for Non-commercial Research Use are permitted, provided that the checkpoints do not contain or substantially reproduce restricted dataset recordings.
Users must not intentionally design or distribute models for the purpose of extracting or reconstructing protected dataset recordings.
Any released checkpoints must comply with applicable third-party rights and the other conditions of these Terms.
Commercial deployment or use of models developed using the dataset under these Terms is not authorized without separate written permission from the relevant rights holders. Publication of model checkpoints does not grant commercial rights to the underlying dataset recordings.
Short research audio demonstrations, including separated vocal excerpts, may be published for Non-commercial Research Use to illustrate research methods or results, subject to applicable third-party rights and any additional permissions required from the relevant rights holders.
Such demonstrations must not substantially reproduce or replace the original dataset recordings or enable their reconstruction.
This permission does not authorize redistribution of full recordings, raw stems, mixtures, or the dataset itself.
6. Attribution
Users must appropriately acknowledge ACappellaSet, identify the dataset version used, and retain the contributor credits supplied with the release.
In publications resulting from use of the dataset, users must cite the official accompanying paper using the citation provided with the dataset, when available.
Until the paper citation becomes available, users must cite the dataset name, version, and repository URL.
Users must not imply endorsement by the dataset creators, performers, or rights holders.
7. Data Security
Users must implement reasonable security measures to prevent unauthorized access to or disclosure of the dataset.
Storage and processing in appropriately secured personal, institutional, or cloud research environments are permitted, provided they comply with these Terms.
8. Access Revocation and Termination
The dataset administrator may suspend or revoke access for violations of these Terms or other legitimate legal or rights-related reasons.
Upon termination, users must cease using the restricted dataset and delete or securely dispose of restricted copies as specified in the termination notice, subject to applicable legal retention requirements.
Revoking platform access does not automatically delete previously downloaded files.
9. Copyright and Third-Party Rights
All rights not expressly granted are reserved.
The dataset contains original compositions, cover songs, and medleys, including musical works and recordings subject to rights held by multiple parties.
These Terms apply only to rights that the dataset providers are authorized to license. They do not transfer ownership or grant additional rights to third-party copyrighted works.
The ACappellaSet Dataset Team does not claim ownership of third-party musical compositions or lyrics.
The release must not include materials for which necessary distribution and research-use permissions have not been established.
10. Updates
Each release of these Terms must have an identifiable version and effective date.
Material changes imposing additional user obligations must be clearly communicated and, where required, accepted by users before continued access.
Updated Terms apply prospectively and do not retroactively modify previously granted permissions unless expressly agreed upon by the affected users.
- Downloads last month
- 51