boobalanit/ComposerMind-OpenAudio-48k

ComposerMind OpenAudio-48k: High-Fidelity Audio Dataset for Machine Learning Research

Overview

ComposerMind OpenAudio-48k is a research-oriented collection of 94 synthetically produced, high-energy electronic music tracks, designed around fast-paced world exploration, interactive gaming, action sequences, and immersive electronic soundscapes.

Each track is accompanied by rich acoustic annotations generated using BeatMind AI (ComposerMind's proprietary acoustic analysis model and analysis pipeline), including tempo (BPM), musical key and Camelot harmonic notation, harmonic/percussive energy characteristics, and audio dynamics.

The dataset also provides a natural-language prompt for each track, intended to support text-to-audio generation, music captioning, semantic retrieval, and audio–text representation learning.

Acoustic Production Provenance:
ComposerMind OpenAudio-48k is a Studio-Grade, In-The-Box (ITB) Electronic Music Corpus. Tracks are composed and digitally synthesized using professional studio sound generators, sequencers, and synthesizers inside a digital audio workstation (DAW). Every track was mastered and audited through WaveWash DSP to ensure zero digital overs ($0\text{ dBFS}$ clamping), balanced dynamic crest factor ($15.1\text{ dB}$ avg), and uniform 48kHz/24-bit headroom.

---

Research Objectives

The dataset is designed to support:

• Music Information Retrieval (MIR): tempo estimation, key detection, harmonic analysis, rhythmic characterization, and audio feature extraction.
• Generative Music AI: prompt-conditioned generation experiments using compatible models and frameworks such as MusicGen, AudioCraft, and Stable Audio Open.
• Audio–Text Representation Learning: matching music recordings with natural-language descriptions for retrieval and embedding research.
• Game Audio Research: exploration of energetic electronic music suitable for interactive worlds, action environments, and gameplay experiences.
• Dataset Quality and Benchmarking: reproducible testing of audio analysis pipelines, metadata consistency, and acoustic feature estimation.

Intended Users

The collection is intended for academic researchers, AI/ML engineers, music information retrieval specialists, game-audio developers, generative music researchers, sound designers, students, and music technology companies.

Annotation Methodology

Acoustic annotations are produced using ComposerMind's proprietary BeatMind AI analysis workflow. The supplied annotations are automatically extracted labels and provide a standardized baseline for benchmark evaluation.

Natural-language prompts are provided as conditioning and descriptive metadata for Text-to-Audio generative models (such as MusicGen, AudioCraft, Stable Audio Open, and MusicLM).

---

Technical Specifications

• Dataset Name: ComposerMind OpenAudio-48k
• Creator & Author: Boobalan Arjunan
• Publisher: ComposerMind Labs
• Copyright: Copyright © 2024 Boobalan Arjunan
• License: Creative Commons Attribution 4.0 International (CC BY 4.0)
• Total Unique Tracks: 94 tracks (OpenAudio48k_Track01.wav through OpenAudio48k_Track94.wav)
• Total Audio Duration: 3.88 hours (13,985.6 seconds / 233.1 minutes)
• Dataset Disk Size: 3.75 GB (uncompressed 24-bit linear PCM)
• Sampling Rate: 48,000 Hz (uniform across 100% of files)
• Bit Depth: 24-bit Linear PCM (pcm_s24le)
• Channels: 2 (Stereo)
• Dynamic Range & Signal Integrity:
  • Peak Level: Maximum -0.01 dBFS, Average -1.98 dBFS
  • Digital Overs / Clipping: 0 / 94 files (clean headroom preserved)
  • Average Dynamic Range (Crest Factor): 15.10 dB (Range: 11.33 dB – 17.90 dB)
  • Average RMS Energy: -17.09 dBFS (balanced commercial game-audio standard)
• Tempo & Key Profile:
  • BPM Range: 93.7 – 176.5 BPM (Average: 123.8 BPM)
  • Key Distribution: Diverse coverage across minor and major keys with exact Camelot codes (6A, 7A, 9A, 12B, etc.)
  • Timbre Breakdown: Average 80.1% Harmonic / 19.9% Percussive energy

---

Dataset Structure

ComposerMind-OpenAudio-48k/
├── metadata.csv                 # Complete 94-track manifest (23 columns)
├── dataset-metadata.json        # Kaggle 10.0 Usability dataset specification
├── starter_eda_pytorch.ipynb    # Ingestion-ready PyTorch Dataset/DataLoader starter
├── LICENSE                      # Creative Commons Attribution 4.0 International
├── README.md                    # Official dataset card & benchmark guide
├── scripts/
│   └── audit_dataset.py         # Non-destructive WAV verification tool
└── docs/
    └── METADATA_REVIEW.md       # Audit history & verification log

Metadata Schema (metadata.csv — 23 Columns)

The reviewed release manifest contains 23 high-relevance attributes:

1. file_name: Standardized audio filename (OpenAudio48k_TrackXX.wav)
2. track_id: Unique track identifier (Track01 - Track94)
3. title: Track title (ComposerMind OpenAudio-48k - Track XX)
4. artist: Creator / Author (Boobalan Arjunan)
5. publisher: Publishing entity (ComposerMind Labs)
6. copyright: Copyright © 2024 Boobalan Arjunan
7. license: CC-BY-4.0
8. primary_genre: Electronic
9. sub_genre: Game Audio / Drum & Bass / Breakbeat
10. mood_label: Energetic / Dynamic Exploration
11. tempo_bpm: Continuous tempo detected via BeatMind AI
12. musical_key: Detected musical key (e.g. G Minor, E Minor)
13. camelot_code: Camelot Wheel harmonic mixing code (6A, 9A, etc.)
14. harmonic_ratio: Percentage of harmonic/tonal energy (0–100%)
15. percussive_ratio: Percentage of transient/percussive energy (0–100%)
16. crest_factor_db: Dynamic range (peak-to-RMS ratio in dB)
17. duration_seconds: Precise duration in seconds
18. sample_rate: Sampling frequency (48,000 Hz)
19. channels: Channel layout (2 = Stereo)
20. bit_depth: Audio resolution (24 bits)
21. peak_dbfs: Maximum absolute peak level in dBFS
22. rms_dbfs: Root-mean-square energy level in dBFS
23. text_prompt: Turnkey natural language training caption for Text-to-Music models

(Note: record_label, musical_scale, bpm_confidence, and groove_quantization were streamlined to remove redundancy and ensure maximum interpretability for research).

---

Python Ingestion Example

Training / Ingestion with PyTorch & Pandas

import os
import pandas as pd
import torchaudio

dataset_dir = "/path/to/ComposerMind_OpenAudio_48k"
df = pd.read_csv(os.path.join(dataset_dir, "metadata.csv"))

# Inspect ground-truth labels
print(df[["file_name", "tempo_bpm", "musical_key", "camelot_code", "text_prompt"]].head(3))

# Load audio waveform
sample_row = df.iloc[0]
audio_path = os.path.join(dataset_dir, sample_row["file_name"])
waveform, sample_rate = torchaudio.load(audio_path)

print(f"Loaded: {sample_row['title']}")
print(f"Tensor Shape: {waveform.shape}, Sample Rate: {sample_rate} Hz")
print(f"Text Prompt: {sample_row['text_prompt']}")

---

Audio Download & Public Registries

The uncompressed 3.75 GB 24-bit audio tracks are hosted on public open-access data registries:

• Hugging Face Hub: huggingface.co/datasets/boobalanit/ComposerMind-OpenAudio-48k
• Kaggle Dataset: kaggle.com/datasets/boobalanarjunan/composermind-openaudio-48k
• Zenodo DOI: doi.org/10.5281/zenodo.24050283

---

Attribution & Citation

Published under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

BibTeX

@dataset{arjunan2026composermind,
  title     = {ComposerMind OpenAudio-48k: Studio Game Audio Benchmark & Generative Music Corpus},
  author    = {Arjunan, Boobalan},
  publisher = {ComposerMind Labs},
  year      = {2026},
  doi       = {10.5281/zenodo.24050283},
  url       = {https://doi.org/10.5281/zenodo.24050283},
  license   = {Creative Commons Attribution 4.0 International (CC BY 4.0)}
}