Datasets:
REDACTED-Tunes Benchmark
Part of the REDACTED-Tunes project — the full 1.4 M-track corpus is available at https://anonymous-hf.up.railway.app/a/pib3syfuxp02/
A 10,521-song perceptual evaluation dataset for AI music detection and quality assessment, anchored in the REDACTED-Tunes corpus. The benchmark pairs AI-generated tracks from seven commercial platforms with human-produced commercial recordings and enriches a subset with 591 human perceptual annotations from 61 participants, enabling rigorous study of authenticity detection, quality perception, and out-of-distribution generalisation for text-to-music systems.
Dataset Summary
| Property | Value |
|---|---|
| Total songs | 10,521 |
| AI-generated | 9,681 (92.0 %) |
| Human recordings | 840 (8.0 %) |
| AI platforms | 7 (Suno, Udio, Mureka, Sonauto, Lyria 3, Riffusion, SilverknightAI) |
| Splits | train / validation / test / test_ood |
| Annotated songs | 572 |
| Human annotation trials | 591 (61 participants) |
| Annotation dimensions | 6 (authenticity + 5 quality ratings, 1–10 scale) |
| License | Apache 2.0 |
Quick Start
# Download the dataset archive via the anonymous HuggingFace mirror
wget "https://anonymous-hf.up.railway.app/api/a/8k5dg0m61yir/download/" -O repo.zip
unzip repo.zip -d anonymous_repo
import datasets
ds = datasets.load_dataset("anonymous_repo")
# Full train split
train = ds["train"]
# Test split — every row has human annotations attached
test = ds["test"]
sample = test[0]
print(sample["title"], sample["source"])
print(sample["annotations"]["authenticity_assessment"]) # list of human labels
print(sample["annotations"]["aesthetic_quality"]) # list of 1–7 ratings
# OOD split — Lyria 3, Riffusion, SilverknightAI + rare human songs
test_ood = ds["test_ood"]
Dataset Construction
Song Pool
The in-distribution (id) AI pool consists of:
- Suno, Udio, and Mureka (~2,500 songs each) — randomly sampled from the REDACTED-Tunes large-scale corpus.
- Sonauto (~2,063 songs) — manually downloaded from publicly accessible Sonauto links and added specifically for this benchmark; these tracks are not part of the REDACTED-Tunes corpus.
Human reference songs are 30-second Apple Music open previews sourced via the Genius metadata index, genre-matched to the AI pool; these are also not part of the REDACTED-Tunes corpus.
The out-of-distribution (OOD) pool introduces platforms absent from REDACTED-Tunes: Lyria 3 (Google DeepMind, 51 songs), Riffusion (51 songs), and SilverknightAI (16 songs), all manually downloaded from publicly accessible links. The OOD human subset (50 songs) was selected to maximise metadata rarity — uncommon languages, genres, and artists not present in the id human pool.
Split Composition
| Split | Total | Human | Suno ★ | Udio ★ | Mureka ★ | Sonauto | OOD sources |
|---|---|---|---|---|---|---|---|
| train | 8,826 | 685 | 2,131 | 2,145 | 2,123 | 1,742 | — |
| validation | 1,005 | 74 | 243 | 242 | 248 | 198 | — |
| test | 522 | 31 | 126 | 113 | 129 | 123 | — |
| test_ood | 168 | 50 | — | — | — | — | Lyria 3 (51), Riffusion (51), SilverknightAI (16) |
| Total | 10,521 | 840 | 2,500 | 2,500 | 2,500 | 2,063 | 118 |
★ Sourced from the REDACTED-Tunes large-scale corpus. All other audio (Human, Sonauto, and OOD sources) was manually downloaded and is exclusive to this benchmark.
Suno model version breakdown: chirp-v3 (1,602), chirp-v4 (502), chirp-auk (210), chirp-bluejay (113), chirp-v2 (34), chirp-chirp (18), chirp-ahi (15), chirp-v1 (5), chirp-bass (1).
Human Annotation Study
Design
Annotations were collected in two phases via a purpose-built web platform:
Phase 1 (v1, 403 trials, collected 2026-04-07): Participants rated AI-generated songs only at full length. The label distribution (all AI) was not disclosed, testing blind AI-detection from learned heuristics. Sources: Suno 105, Mureka 103, Sonauto 103, Udio 92.
Phase 2 (v2-only additions, 188 trials, collected 2026-04-20): A 50/50 balanced design introduced human recordings alongside AI songs. All clips were trimmed to 30 seconds (human songs: fixed Apple Music preview; AI songs: uniformly random crop). Sources: human 84, Mureka 30, Suno 28, Udio 24, Sonauto 22.
Each trial asked the participant to:
- Judge authenticity:
real/ai-generated/uncertain - Rate familiarity:
never/heard_before/know_well - Rate five perceptual dimensions on a continuous 1–10 slider: aesthetic quality, production quality, emotional engagement, musical creativity, playlist likelihood
- Optionally tag AI aspects, mood, and write free-text comments
Participants
| Property | Value |
|---|---|
| Registered | 91 |
| Active (≥1 annotation) | 61 |
| Mean age | 29.4 yrs (exact age reported by 32/61) |
| Median age | 27.0 yrs |
| Mean formal training | 4.0 yrs (median 0) |
| Annotations per participant | mean 9.7 ± 10.7 (min 1, max 50) |
Musical engagement (participant-level): enthusiast 20, casual 19, musician 16, professional 6.
AI music experience: use occasionally 19, tried once or twice 15, professional experience 9, use regularly 8, heard about but never tried 5, never used 5.
Listening setup: 82% used headphones or earbuds (over-ear 19, earbuds 15, on-ear 11); 82% listened in a quiet room.
Top self-reported genres: Rock, Pop, Electronic, Indie, Classical, Alternative, Hip-Hop, Metal, Jazz, Punk.
Results
Authenticity Detection Accuracy
| Source | Trials | Correct | Accuracy |
|---|---|---|---|
| Human | 84 | 54 | 64.3 % |
| Suno | 133 | 74 | 55.6 % |
| Mureka | 133 | 76 | 57.1 % |
| Sonauto | 125 | 69 | 55.2 % |
| Udio | 116 | 48 | 41.4 % |
Overall authenticity breakdown: ai-generated 289 (48.9 %), real 225 (38.1 %), uncertain 77 (13.0 %).
Perceptual Quality Ratings (1–10 scale, mean ± σ)
| Source | Aesthetic | Production | Emotional | Creativity | Playlist |
|---|---|---|---|---|---|
| Human | 5.68 ± 1.94 | 6.06 ± 1.69 | 5.02 ± 1.91 | 5.42 ± 2.02 | 4.31 ± 2.02 |
| Suno | 5.18 ± 2.31 | 5.34 ± 2.34 | 4.79 ± 2.21 | 4.70 ± 2.18 | 3.99 ± 2.42 |
| Udio | 5.15 ± 1.98 | 5.21 ± 2.10 | 4.78 ± 2.06 | 4.83 ± 2.10 | 3.85 ± 2.15 |
| Mureka | 5.02 ± 2.22 | 5.19 ± 2.12 | 4.32 ± 2.13 | 4.36 ± 2.09 | 3.45 ± 2.20 |
| Sonauto | 4.94 ± 2.38 | 5.18 ± 2.54 | 4.56 ± 2.30 | 4.81 ± 2.25 | 3.69 ± 2.37 |
Mean annotation duration: 129.3 s (median 80.8 s, σ = 224.8 s).
Key finding: Human songs score higher on all five quality dimensions, yet human-detection accuracy is only 64.3 % — a quality-authenticity halo effect where higher quality does not reliably predict a "real" verdict. Udio's near-chance detection rate (41.4 %) indicates perceptual near-parity with human music on the production quality axis.
Schema
Each row represents one song. All fields are present in all splits; annotations is an empty list for train and validation rows.
Song-level fields
| Field | Type | Description |
|---|---|---|
uuid |
string | UUID of the song |
song_id |
string | Prefixed identifier, e.g. suno_<uuid> or human_<uuid> |
source |
string | Platform: human, suno, udio, mureka, sonauto, lyria3, riffusion, silverknightai |
label |
string | Binary label: human or ai |
split_pool |
string | id (in-distribution) or ood |
audio_url |
string | Direct audio URL (CDN; may expire) |
title |
string | Song title |
model |
string | Model version where available (e.g. chirp-v4 for Suno) |
duration_ms |
int32 | Duration in milliseconds (null if unavailable) |
meta_json |
string | Source-specific metadata serialised as a JSON string |
is_held_out |
bool | Whether this song was held out from the main corpus |
ood_reason |
string | Reason for OOD assignment (human_metadata_rarity or empty) |
annotations |
struct (list) | Zero or more human annotation records (see below) |
Annotation sub-fields (per annotation in the list)
| Field | Type | Description |
|---|---|---|
annotation_id |
string | UUID of the annotation trial |
session_id |
string | UUID of the annotation session |
participant_id |
string | Pseudonymised participant UUID |
authenticity_assessment |
string | real / ai-generated / uncertain |
familiarity_level |
string | never / familiar / know / uncertain |
aesthetic_quality |
float32 | Continuous slider 1.0–10.0 (1 = Poor, 10 = Excellent) |
production_quality |
float32 | Continuous slider 1.0–10.0 (1 = Not at all, 10 = Very much) |
emotional_engagement |
float32 | Continuous slider 1.0–10.0 (1 = Not at all, 10 = Very much) |
musical_creativity |
float32 | Continuous slider 1.0–10.0 (1 = Not at all, 10 = Very much) |
playlist_likelihood |
float32 | Continuous slider 1.0–10.0 (1 = Not at all, 10 = Very much) |
ai_aspects |
string | Free-text: which aspects sounded AI-generated |
mood_tags |
string | JSON array of mood tags selected by the participant |
aesthetic_comment |
string | Free-text aesthetic comment |
song_description |
string | Free-text song description |
annotation_duration_ms |
int32 | Time spent on this annotation trial |
created_at |
string | ISO 8601 timestamp |
participant_age |
int16 | Exact age (null if not provided) |
participant_age_range |
string | Age bracket (e.g. 25-34; null if exact age provided) |
participant_musical_genres |
string | JSON array of self-reported genres |
participant_musical_engagement |
string | casual / enthusiast / musician / professional |
participant_formal_training_years |
int16 | Years of formal music training |
participant_listening_device |
string | Headphones (over-ear), Earbuds/In-ear, etc. |
participant_listening_context |
string | alone / with others |
participant_environment |
string | Quiet room, Office/workplace, etc. |
participant_ai_music_experience |
string | Self-reported AI music familiarity |
Limitations
- Annotation coverage is partial: 591 trials cover 572 / 10,521 songs (~5.4 %); all annotated songs are in the
testortest_oodsplit. - Phase 1 (v1) exposed participants only to AI music, which may have calibrated their detectors toward platform artefacts rather than general authenticity cues.
- Audio clips in phase 2 are trimmed to 30 seconds; full-song judgements may differ.
- The participant pool (n = 61, mean age 29.4) is small and skews toward regular or occasional AI-music users.
- OOD sources (Lyria 3, Riffusion, SilverknightAI) carry no human annotations.
- CDN audio URLs point to third-party servers and may expire without notice.
- The benchmark reflects model capabilities as of April 2026.
Ethical Considerations
Human songs are 30-second preview clips served from publicly accessible Apple Music links; no full tracks are distributed. AI-generated songs are served from platform CDNs under their respective public access terms. No participant personal data beyond pseudonymised demographics is released. All participants provided informed consent for anonymous open data sharing.
Ethics
Participant Privacy and Consent
All annotation participants were recruited voluntarily and gave informed consent prior to the study. Each participant was assigned an opaque UUID; no name, email address, or other directly identifying information is stored in or released with the dataset. Demographic data (age, musical background, listening device, etc.) is released only in aggregate form within each annotation record and is not linkable to any external identity.
Audio Content Rights
Human reference tracks are distributed exclusively as 30-second Apple Music open previews, consistent with the terms under which Apple Music exposes these excerpts publicly. No full audio tracks are included. AI-generated tracks are served via public CDN links from the originating platforms (Suno, Udio, Mureka, Sonauto, Lyria 3, Riffusion, SilverknightAI); their availability is subject to each platform's terms of service and may change over time.
Intended and Unintended Uses
This dataset is intended for academic research on AI music detection, perceptual quality modelling, and human–AI comparative evaluation. It is not intended to be used to build commercial AI-music detection products without further validation, to deanonymise participants or artists, or to infer personally sensitive attributes from audio features.
Bias and Fairness
The human reference pool is drawn from commercially indexed music (Apple Music via Genius), which over-represents mainstream Western genres (Rock, Pop, Electronic). The AI generator pool reflects the four platforms most prevalent on the open web as of early 2026; less commercially prominent generators are relegated to the OOD split. The participant pool (n = 61, mean age 29.4) skews toward young, music-engaged, and AI-familiar individuals, and cannot be assumed to generalise to broader listener populations. Researchers should account for these selection biases when drawing conclusions about general human perception.
Dual-Use Considerations
Benchmark results that reveal the most perceptually convincing AI generators could inadvertently assist actors seeking to produce undetectable synthetic media. We release this dataset to support the development of robust detection methods and transparent model evaluation, and we encourage the community to weigh potential misuse when building systems trained on these annotations.
Environmental Impact
No new audio was generated for this benchmark. All AI tracks were collected from publicly accessible platform outputs; no additional model training or large-scale inference was conducted by the dataset curators.
Citation
@dataset{redacted_tunes_benchmark_2026,
title = {{REDACTED}-Tunes-Benchmark: A Perceptual Benchmark for AI-Generated Music},
author = {REDACTED},
year = {2026},
url = {https://anonymous-hf.up.railway.app/a/8k5dg0m61yir/},
license = {Apache-2.0},
}
Related Resources
- Full corpus (REDACTED-Tunes, 1.4 M tracks): https://anonymous-hf.up.railway.app/a/pib3syfuxp02/
- Croissant metadata:
hf_dataset/croissant.json(MLCommons Croissant 1.1, includes all six RAI fields)
- Downloads last month
- 62