LAION-Tunes

908,174 AI-generated music tracks from 5 platforms, annotated with captions, transcriptions, embeddings, aesthetics scores, and NSFW safety labels. Includes a full-text and vector search engine with a web UI.

Quick Stats

Metric	Value
Total tracks	908,174
Subsets	Mureka (383K), Suno (308K), Udio (115K), Riffusion (99K), Sonauto (3K)
Has caption (Music-Whisper)	832,944 (91.7%)
Has transcription (Parakeet ASR)	514,203 (56.6%)
Instrumental	394,246 (43.4%)
NSFW flagged (very likely + likely)	12,860 (1.4%)

Dataset Description

LAION-Tunes is a curated metadata and annotation dataset derived from ai-music/ai-music-deduplicated, a collection of publicly available AI-generated music from Suno, Udio, Mureka, Riffusion, and Sonauto.

This dataset does NOT contain audio files. It contains metadata, annotations, embeddings, and search indices. Audio URLs pointing to the original hosting platforms are included for reference.

What's Included

For each track:

• Metadata: title, tags, genre, mood, duration, play count, upvote count
• Music-Whisper Caption: AI-generated music description using laion/music-whisper (fine-tuned OpenAI Whisper Small)
• Parakeet ASR Transcription: Speech-to-text using nvidia/parakeet-tdt-0.6b-v3 with word-level timestamps
• Sentence Embeddings: 768-dim embeddings via google/embeddinggemma-300m for tags, captions, transcriptions, lyrics, and moods
• Whisper Audio Embeddings: 768-dim mean-pooled encoder embeddings from Music-Whisper for audio similarity search
• Aesthetics Scores: coherence, musicality, memorability, clarity, naturalness (computed from Music-Whisper)
• NSFW Safety Labels: three-tier classification (very_likely_nsfw / likely_nsfw / likely_sfw) across gore, sexual, and hate speech dimensions
• Pre-built Search Indices: FAISS vector indices and BM25 text indices ready to serve

Annotation Pipeline

The annotation pipeline processes the original TAR files from ai-music-deduplicated:

1. Music-Whisper (laion/music-whisper): Generates music captions describing instruments, genre, mood, tempo, etc.
2. Parakeet TDT 0.6B (nvidia/parakeet-tdt-0.6b-v3): ASR transcription with word-level timestamps for vocal content
3. EmbeddingGemma 300M (google/embeddinggemma-300m): Computes 768-dim sentence embeddings for captions, transcriptions, tags, lyrics, moods
4. Whisper Encoder Embeddings: Mean-pooled encoder hidden states from Music-Whisper for audio fingerprinting/similarity
5. NSFW Classification: Cosine similarity of transcription embeddings against reference prompts for gore/violence, sexual content, and hate speech

Repository Structure

laion-tunes/
├── README.md                          # This file
├── server.py                          # FastAPI search server (main application)
├── index.html                         # Web UI (dark-mode, single-page app)
├── build_search_index.py              # Index builder script
├── update_indices.py                  # Incremental index updater
├── migrate_add_language.py            # Language detection migration
├── nsfw_safety_report.html            # Interactive NSFW analysis report
├── nsfw_analysis_data.json            # Raw NSFW analysis data
│
├── public/                            # Annotated metadata parquets (8.3 GB)
│   ├── mureka_000000.tar.parquet      # One parquet per source TAR file
│   ├── mureka_000001.tar.parquet
│   ├── ...
│   └── udio_000015.tar.parquet        # 159 parquet files total
│
├── search_index/                      # Pre-built search indices (16 GB)
│   ├── metadata.db                    # SQLite database (908K tracks, 2.7 GB)
│   ├── faiss_tag.index                # FAISS IndexFlatIP - tag embeddings (2.6 GB)
│   ├── faiss_whisper.index            # FAISS IndexFlatIP - audio embeddings (2.6 GB)
│   ├── faiss_caption.index            # FAISS IndexFlatIP - caption embeddings (2.3 GB)
│   ├── faiss_transcription.index      # FAISS IndexFlatIP - transcription embeddings (1.5 GB)
│   ├── faiss_lyric.index              # FAISS IndexFlatIP - lyric embeddings (1.4 GB)
│   ├── faiss_mood.index               # FAISS IndexFlatIP - mood embeddings (1.1 GB)
│   ├── idmap_*.npy                    # Row ID mappings for each FAISS index
│   ├── bm25_tags.pkl                  # BM25 text index for tags (114 MB)
│   ├── bm25_caption.pkl               # BM25 text index for captions (609 MB)
│   └── bm25_transcription.pkl         # BM25 text index for transcriptions (392 MB)
│
└── whisper_embeddings/                # Raw Whisper encoder embeddings (1.6 GB)
    ├── mureka_000000.tar.npz          # One NPZ per source TAR file
    ├── ...
    └── udio_000015.tar.npz            # 159 NPZ files total

Data Format

Parquet Files (public/)

Each parquet file corresponds to one TAR file from the source dataset and contains these columns:

Column	Type	Description
filename	str	Filename within the source TAR
tar_file	str	Source TAR filename
audio_url	str	Original audio URL (mp3/m4a/ogg)
subset	str	Source platform (suno/udio/mureka/riffusion/sonauto)
title	str	Track title
tags	str	Comma-separated genre/style tags
mood	str	Mood tags
lyrics	str	Lyrics (if available, from source metadata)
duration_seconds	float	Track duration
play_count	int	Play count on source platform
upvote_count	int	Like/upvote count
music_whisper_caption	str	Music-Whisper generated caption
parakeet_transcription	str	Parakeet ASR transcription (plain text)
parakeet_transcription_with_timestamps	str	ASR with word-level timestamps
tag_embedding	list[float]	768-dim EmbeddingGemma embedding of tags
caption_embedding	list[float]	768-dim EmbeddingGemma embedding of caption
transcription_embedding	list[float]	768-dim EmbeddingGemma embedding of transcription
lyric_embedding	list[float]	768-dim EmbeddingGemma embedding of lyrics
mood_embedding	list[float]	768-dim EmbeddingGemma embedding of mood

Whisper Embeddings (whisper_embeddings/)

NPZ files containing mean-pooled Whisper encoder hidden states:

• embeddings: float32 array of shape (N, 768) - L2-normalized
• filenames: string array of filenames matching the parquet entries

SQLite Database (search_index/metadata.db)

The tracks table contains all 908,174 tracks with 34 columns including metadata, aesthetics scores, annotation flags, and NSFW safety labels. The row_id column is the primary key used by all FAISS indices.

FAISS Indices (search_index/faiss_*.index)

All indices are IndexFlatIP (inner product / cosine similarity for L2-normalized vectors) with 768 dimensions. Each index has a corresponding idmap_*.npy that maps FAISS internal indices to SQLite row_id values.

Index	Vectors	Description
faiss_tag	908,241	Tag text embeddings
faiss_whisper	908,174	Audio encoder embeddings (music similarity)
faiss_caption	798,858	Music-Whisper caption embeddings
faiss_transcription	511,610	ASR transcription embeddings
faiss_lyric	479,313	Lyrics embeddings
faiss_mood	383,616	Mood text embeddings

NSFW Safety Labels

Each track has NSFW safety scores and labels across three dimensions:

Dimension	Strict Threshold	Moderate Threshold	Very Likely NSFW	Likely NSFW
Gore/Violence	>= 0.3779	>= 0.3540	2,437 (0.27%)	2,293 (0.25%)
Sexual Content	>= 0.3584	>= 0.3234	3,367 (0.37%)	2,689 (0.30%)
Hate Speech	>= 0.3633	>= 0.3382	2,786 (0.31%)	2,316 (0.26%)
Overall (conservative)	-	-	6,762 (0.74%)	6,098 (0.67%)

• very_likely_nsfw: cosine similarity above strict threshold
• likely_nsfw: between strict and moderate thresholds
• likely_sfw: below moderate threshold
• nsfw_overall_label: conservative (worst label across all three dimensions wins)

The raw cosine similarity scores (nsfw_gore_sim, nsfw_sexual_sim, nsfw_hate_sim) are stored so you can apply your own thresholds.

Running the Search Server

Prerequisites

pip install fastapi uvicorn faiss-cpu numpy pandas sentence-transformers torch scipy tqdm

Option 1: With HF Text Embeddings Inference (Recommended)

TEI provides fast CPU-based embedding serving (~25ms per query vs ~430ms with Python):

# Start TEI (requires Docker)
docker run -d --name tei-embeddings \
  -p 8090:80 \
  -e HF_TOKEN=your_token \
  ghcr.io/huggingface/text-embeddings-inference:cpu-latest \
  --model-id google/embeddinggemma-300m \
  --max-batch-requests 4

# Start the server with TEI backend
python server.py --port 7860 --gpu 0 --tei-url http://localhost:8090

Option 2: Direct Python Inference

python server.py --port 7860 --gpu 0

This loads EmbeddingGemma 300M and Music-Whisper encoder into Python directly.

Server Flags

Flag	Default	Description
--port	7860	HTTP port
--host	0.0.0.0	Bind address
--gpu	0	GPU ID for Whisper encoder
--tei-url	None	TEI server URL for text embeddings
--no-whisper	False	Skip loading Whisper encoder (disables audio similarity search)

What Loads at Startup

1. 6 FAISS indices (tag, whisper, caption, transcription, lyric, mood)
2. 3 BM25 indices (tags, caption, transcription)
3. SQLite database (908K tracks)
4. Text embedder: TEI backend or Python SentenceTransformer
5. Music-Whisper encoder (on GPU): for audio upload similarity search

Total memory: ~20 GB RAM + ~200 MB GPU VRAM

Search API

POST /api/search

Main search endpoint combining vector search, BM25, and metadata filtering.

{
  "query": "upbeat electronic dance music",
  "search_fields": ["tag", "caption"],
  "bm25_fields": ["tags", "caption"],
  "top_k": 50,
  "bm25_weight": 0.3,
  "min_score": 4.0,
  "max_score": 10.0,
  "subsets": ["suno", "udio"],
  "nsfw_filter": "sfw_only",
  "stage2_enabled": true,
  "stage2_field": "caption",
  "stage2_top_k": 200
}

Search fields (FAISS vector search):

• tag: search by music tags/genres
• caption: search by Music-Whisper captions
• transcription: search by ASR transcriptions
• lyric: search by lyrics content
• mood: search by mood descriptors
• whisper: search by audio embedding similarity

BM25 fields (text search):

• tags, caption, transcription

Filters:

• min_score / max_score: aesthetics score range (0-10)
• subsets: list of source platforms
• nsfw_filter: "sfw_only" or "nsfw_only" or null for all

Two-stage search: First retrieves stage2_top_k candidates, then re-ranks by a second field.

POST /api/search_by_audio

Upload an audio file to find similar tracks by audio fingerprint.

curl -X POST http://localhost:7860/api/search_by_audio \
  -F "file=@song.mp3" \
  -F "top_k=20" \
  -F "nsfw_filter=sfw_only"

POST /api/search_similar

Find tracks similar to an existing track by row_id.

{
  "row_id": 12345,
  "field": "whisper",
  "top_k": 20
}

GET /api/stats

Returns dataset statistics (total tracks, index sizes, NSFW counts).

Building the Index from Scratch

If you want to rebuild the search indices from the parquet files:

python build_search_index.py --force

This reads all parquets from public/ (and optionally private/ for additional embeddings), builds the SQLite database, FAISS indices, and BM25 indices. Takes ~30 minutes on a modern machine.

Source Data

The original audio data comes from ai-music/ai-music-deduplicated, organized by platform:

Platform	Tracks	Audio Format	Notable Fields
Mureka	383,549	MP3	genres, moods, model version
Suno	307,539	MP3	tags (in metadata), prompt, model_name, explicit flag
Udio	115,140	MP3	tags (array), lyrics, prompt, likes, plays
Riffusion	99,228	M4A	sound (style description), lyrics_timestamped, conditions
Sonauto	2,718	OGG	tags (array), description, keyword

Models Used

Model	Purpose	Output
laion/music-whisper	Music captioning + audio embeddings	Text caption + 768-dim encoder embedding
nvidia/parakeet-tdt-0.6b-v3	ASR transcription	Text + word-level timestamps
google/embeddinggemma-300m	Text sentence embeddings	768-dim L2-normalized vectors

License

Apache 2.0

Citation

@misc{laion-tunes-2025,
  title={LAION-Tunes: Annotated AI Music Search Dataset},
  author={LAION},
  year={2025},
  url={https://huggingface.co/datasets/laion/laion-tunes}
}