--- license: other task_categories: - audio-classification - question-answering - text-generation language: - en - ko pretty_name: VocalCoachBench size_categories: - 10K **Release version: `canonical515_physical_audio_1064refs`.** > Canonical Top-3 denominator: **515 recordings (A=207, B=308)**. > `top3_issue_annotations.jsonl` contains 1,064 annotations covering all 515 > recording IDs. An earlier filename-only join covered 511 IDs because four > pairs of distinct physical audio files shared a submission filename. That > pre-repair 511-ID count must not be used as the benchmark denominator. See > `metadata/CANONICAL_EVALUATION.md` and `metadata/top3_join_policy.json`. This release contains 1,064 recording-level expert reviews from 18 vocal trainers: 504 recordings have two experts, nine have six, and two have one. Its 4,506 feedback blocks yield 12,216 atomic claims: 6,162 diagnoses, 4,593 corrections, and 1,461 strengths. Earlier arXiv and dataset snapshots contain counts from before the physical-audio join repairs; the versioned counts and [release notes](metadata/PUBLIC_RELEASE.md) govern this release. Subsets: - `same_song` / Dataset A: 207 *Amazing Grace* performances from DAMP-S-AG. Audio is not redistributed; use `audio_filename` to match the official release. - `diverse_song` / Dataset B: 308 recordings from multiple public singing datasets. Redistributable audio is included when source terms allow it. ## What Is Included | File | Rows | Description | |---|---:|---| | `annotations/recordings.jsonl` | 515 | One row per benchmark recording, with subset/source metadata and audio access fields. | | `annotations/feedback_blocks.jsonl` | 4,506 | Expert coaching feedback blocks, grouped by recording and anonymized annotator; covers all 515 recordings. | | `annotations/atomic_claims.jsonl` | 12,216 | Atomic issue, correction, and strength claims extracted from feedback blocks; covers all 515 recordings. | | `annotations/top3_issue_annotations.jsonl` | 1,064 | Expert Top-3 issue annotations joined by physical-audio identity; covers all 515 recordings. | | `annotations/triplet_rankings.jsonl` | 450 | Raw same-song triplet ranking annotations, with benchmark inclusion fields. | | `annotations/triplet_ranking_eval_units.jsonl` | 219 | One row per unique triplet, including the 2-of-3 pairwise agreement filter. | | `annotations/segment_annotations.jsonl` | 2,526 | Processed local segment feedback annotations over 304 diverse-song recordings; see the version limitation below. | | `annotations/segment_consensus_events.jsonl` | 262 | Segment-conditioned consensus events with IoU >= 0.3 and at least one shared issue tag. | | `annotations/source_datasets.jsonl` | 10 | Source dataset counts, links, licenses, and redistribution status. | | `metadata/category_ontology.json` | - | Seven issue labels and their parent categories. | | `metadata/triplet_ranking_policy.json` | - | Machine-readable triplet ranking inclusion policy and counts. | | `metadata/top3_join_policy.json` | - | Machine-readable canonical Top-3 join policy, counts, source hashes, and fallback audit. | | `metadata/claims_join_policy.json` | - | Physical-audio join policy, counts, and source hashes for feedback blocks and atomic claims. | | `metadata/canonical_top3_recording_inventory.jsonl` | 515 | One-row-per-recording Top-3 coverage audit with annotation counts and human-pair eligibility. | | `metadata/CANONICAL_EVALUATION.md` | - | Frozen canonical Top-3 denominator and evaluation rules. | | `metadata/PUBLIC_RELEASE.md` | - | Release version, final leaderboard definitions, and segment limitations. | | `metadata/excluded_submission_files.json` | - | Historical/non-Top-3 exclusion metadata; superseded for canonical Top-3 by the hash-aware join policy. | | `metadata/release_stats.json` | - | Release-level counts. | | `metadata/croissant_rai_vocalcoachbench.json` | - | Machine-readable schema, source provenance, and Responsible AI metadata. | | `evaluation/leaderboard.json` | - | Versioned final leaderboard with metric definitions, coverage, and provenance. | | `evaluation/README.md` | - | Result definitions, evaluation policies, prompt snapshots, and reproducibility scope. | Audio, when redistributed, is stored in `audio/diverse_song/` and `audio/segment_clips_iou03/`. Rows without redistributed audio have `audio_path` or `clip_path` set to `null`. ## Dataset Browser A lightweight static browser is included at [`demo/index.html`](demo/index.html) for inspecting recordings, redistributed audio, expert feedback blocks, atomic claims, Top-3 issue annotations, consensus segment events, and triplet ranking metadata. ## Source Datasets VocalCoachBench uses existing public singing datasets as audio sources. The benchmark contribution is the vocal-coaching annotation layer, task definitions, evaluation metadata, and code for using these public sources. We do not claim the original public audio datasets as part of our submission. | Source dataset | Records | Audio in this package | License / terms note | Official source | |---|---:|---|---|---| | DAMP-S-AG | 207 | Not included; use official source | Smule Research Data License Agreement | [link](https://zenodo.org/records/3596940) | | PopBuTFy | 85 | Not included; permission pending | CC BY-NC-SA 4.0; author registration terms request that download links not be redistributed | [link](https://github.com/MoonInTheRiver/NeuralSVB) | | N20EMv2 | 65 | Included | CC BY-SA 4.0 | [link](https://zenodo.org/records/10814703) | | LM-SSD | 7 | Included | CC BY 4.0 | [link](https://audiolabs-erlangen.de/resources/MIR/LM-SVR) | | NUS-48E | 41 | Included | CC BY 4.0 | [link](https://zenodo.org/records/19595152) | | MedleyDB | 32 | Not included; permission pending | CC BY-NC-SA 4.0 with author request not to republish without consent | [link](https://medleydb.weebly.com/downloads.html) | | MedleyVox | 12 | Included | CC BY-NC-SA 4.0 | [link](https://zenodo.org/records/7984549) | | MRSSing | 43 | Included | CC BY 4.0 | [link](https://huggingface.co/datasets/verstar/MRSAudio) | | URSing | 14 | Included | CC BY 4.0 | [link](https://zenodo.org/records/6404999) | | GTSinger | 9 | Included | CC BY-NC-SA 4.0 | [link](https://huggingface.co/datasets/GTSinger/GTSinger) | ## Schema Notes The public join key is `recording_id`. For construction of the canonical Dataset-B Top-3, feedback-block, and atomic-claim tables, raw submissions are first resolved to physical recordings by audio SHA-256 and only then assigned `recording_id`; joining raw annotations by `audio_filename` alone is prohibited because filenames are not unique. Most rows also include `dataset` and `subset`; annotator IDs are stable anonymized IDs. Feedback blocks and atomic claims can also be joined by `block_uid`. The final evaluation prompts use these seven issue labels: - `PITCH` - `RHYTHM` - `DICTION` - `BREATH` - `PHONATION` - `TECHNIQUE` - `EXPRESSION` Parent categories are `Musical Accuracy`, `Technical Production`, and `Delivery`. The released annotation files and `metadata/category_ontology.json` retain the legacy label `VOCALIZATION`. It is an alias of `PHONATION`, not an eighth class; normalize the two labels to the same class when scoring. See `metadata/category_ontology.json` for the parent mapping. Triplet ranking evaluation uses only triplets with at least 2-of-3 pairwise agreement between two expert rankings. Use `triplet_ranking_eval_units.jsonl` or filter `triplet_rankings.jsonl` with `benchmark_included == true`. This yields 378 reference annotations over 189 unique evaluation triplets. ## Evaluation and Version Notes The [final leaderboard](evaluation/leaderboard.json), updated 2026-10-26, uses schema `vcb_final_leaderboard_v1` and dataset version `canonical515_physical_audio_1064refs`. Its policy is `strict_all_attempts`: stored first responses are evaluated, and malformed or missing outputs count as errors. Expert references have equal total weight per recording. | Evaluation | Main denominator and interpretation | |---|---| | Top-3 | Seven-class macro-F1 over 515 recordings / 1,064 expert references; each recording has unit total weight. | | Segment | Seven-class macro-F1 over the frozen 262 consensus events. | | Direct and score-derived ranking | 189 triplets / 567 pair decisions, scored against both expert orders: 1,134 reference attempts. These are separate tasks. | | Diagnosis strict hit / hit | 6,162 expert issue claims; strict hit counts `STRICT_HIT`, and hit also includes `COARSE_HIT`. | | Prescription validity / diagnosis-to-prescription E2E | 2,575 fixed slots (515 × 5); missing prescriptions count as failures. | | Human expert-vs-expert Top-3 | 513 recordings with at least two experts; the two single-expert recordings are excluded. | The final leaderboard's Claims extension reports format failures: invalid JSON produces zero emitted claims and remains in the fixed denominators. Main and Simple prompt protocols and concatenated-audio Direct variants are labeled separately. Unavailable results are `null`, not zero. See [evaluation/README.md](evaluation/README.md) for the full metric definitions. **Segment limitation:** the released 2,526 segment annotations cover 304 Dataset-B recordings, and the leaderboard retains its frozen 262-event set. A separately staged physical-audio join repair produces 2,583 annotations over 308 recordings and 271 consensus events. That staged revision is not installed in this release; it needs new audio clips and model predictions before a new benchmark version can replace the 262-event results. See [metadata/PUBLIC_RELEASE.md](metadata/PUBLIC_RELEASE.md). ## Audio Redistribution - DAMP-S-AG same-song audio is excluded. Download it from the official Zenodo release and match by `audio_filename`. - PopBuTFy and MedleyDB audio are excluded pending redistribution permission. - Other diverse-song audio is included when allowed by source terms. Redistributed full-recording audio files are copied from their public source datasets without additional audio processing. The segment clips are derived time crops for the segment-conditioned task. We do not claim copyright or ownership over the original source audio; users should follow and cite the original source datasets and their licenses or terms. If any original dataset author or maintainer requests removal of redistributed audio, or prefers source-link-only access, we will remove the affected audio files from this package and switch those rows to official-source-link based distribution. ## Privacy Annotator names, raw local paths, and source submission filenames containing annotator names are not included in public annotation rows. ## Loading Example ```python from datasets import load_dataset repo_id = "Hayeonbang/VocalCoachBench" recordings = load_dataset(repo_id, "recordings", split="train") claims = load_dataset(repo_id, "atomic_claims", split="train") triplet_units = load_dataset(repo_id, "triplet_ranking_eval_units", split="train") same_song = recordings.filter(lambda row: row["subset"] == "same_song") eval_triplets = triplet_units.filter(lambda row: row["benchmark_included"]) ``` The `train` split name is a storage convention for the annotation tables, not a benchmark training partition. Record the dataset revision when reporting results, and pass `revision=""` to reproduce that snapshot. ## Citation If you use VocalCoachBench, please cite the paper and the original audio datasets relevant to your work: ```bibtex @misc{bang2026vocalcoachbench, title = {VocalCoachBench: Benchmarking Audio-Language Models on Expert Feedback for Singing}, author = {Hayeon Bang and Hounsu Kim and Wonil Kim and Juhan Nam}, year = {2026}, eprint = {2609.04241}, archivePrefix = {arXiv}, url = {https://arxiv.org/abs/2609.04241}, note = {Accepted to the NeurIPS 2026 Evaluations \& Datasets Track} } ```