kroganCodes/sonata-classical-music-dataset

🎹 SONATA (Augmented MAESTRO)

Symmetrically Orchestrated Notes and Audio for Transformer Architectures

The Next-Generation Cross-Modal Benchmark for Classical Music Era Classification & Audio-Symbolic Transformers

---

🌟 Executive Pitch: The Next Big Thing in Music AI & Classification

For nearly a decade, Google Magenta’s MAESTRO (MIDI and Audio Edited for Synchronous TRacks and Organization) has stood as the undisputed gold standard for machine learning in expressive piano performance. Yet for macro-temporal semantic tasks—most notably Musical Era Classification—MAESTRO suffers from an insidious architectural vulnerability: an extreme 65.3% Romantic monoculture.

When modern deep learning models (Convolutional Neural Networks, Variational Autoencoders, Audio Spectrogram Transformers, and Perceiver architectures) are trained on standard benchmarks, they do not learn the contrapuntal, harmonic, or morphological syntax of Western art music. Instead, they collapse into a trivial majority heuristic, overwhelmed by the loss gradients of Chopin, Liszt, and Rachmaninoff, while starving minority eras (Impressionism at 2.8% and Modernism at 2.8%).

🚀 Enter SONATA (Augmented MAESTRO)

SONATA breaks the archive crisis. By fusing Google Magenta MAESTRO v3.0.0, the ASAP Dataset (ISMIR 2020), the ADL Piano MIDI Dataset (ISMIR 2019), and curated master archives from the Classical MIDI Archive, SONATA establishes the first mathematically symmetrical, acoustic-grounded, multi-modal classical era benchmark in MIR history.

           [ BASELINE MAESTRO v3.0.0 ]                    [ SONATA (Augmented MAESTRO) ]
             Severe 65.3% Romantic Skew                     Perfect 16.67% Class Symmetry
 
    Romantic      ████████████████████  65.3%      Baroque       ████████████  200 (16.67%)
    Classical     ████                  14.6%      Classical     ████████████  200 (16.67%)
    Baroque       ████                  12.4%      Romantic      ████████████  200 (16.67%)
    Late Romantic █                      4.3%      Late Romantic ████████████  200 (16.67%)
    Impressionist ▏                      2.8%      Impressionist ████████████  200 (16.67%)
    Modern        ▏                      2.8%      Modern        ████████████  200 (16.67%)
                  (Gini Skew = 0.58)                               (Gini Skew = 0.00)

---

💎 The Four Pillars of Dataset Fusion

SONATA unites the world's most rigorous symbolic transcriptions and concert captures into one cohesive, studio-grade standard:

1. Google Magenta MAESTRO (v3.0.0):
   • High-resolution Disklavier concert recordings from the Minnesota International Piano-e-Competition.
   • Provides all baseline acoustic ground truth and stratified anchor pieces for Classical and Romantic epochs.
2. ASAP Dataset (Accelerate, Score, Alignment, and Piano - ISMIR 2020):
   • Synchronized score-to-performance Disklavier human captures (e.g., Prokofiev Toccata, Op. 11), contributing authentic human micro-timing and polyphonic dynamics.
3. ADL Piano MIDI Dataset (ISMIR 2019):
   • Curated classical pedagogical corpus with vetted tempo maps, polyphonic note matrices, and pristine key signatures.
4. Classical MIDI Master Archives (John Sankey, Steven Ritchie, Jeff Segor):
   • Scholarly sequenced realizations of Scarlatti, Debussy, Ravel, Satie, Fauré, Scriabin, and Bartók, meticulously cataloged to eliminate arrangement artifacts.

---

🔬 Symmetrical Distribution & Multi-Modal Scale

SONATA provides exact 1:1:1:1:1:1 mathematical parity across all six defining epochs of Western art music, completely neutralizing class imbalance gradient bias.

Musical Era	Total Tracks	Total WAVs	Total MIDIs	Audio Duration	Discrete Note Events	Mean RMS Energy	Train (80%)	Val (10%)	Test (10%)
Baroque	200	200	200	14.82 hrs	521,409	0.0612	160	20	20
Classical	200	200	200	26.14 hrs	874,130	0.0594	160	20	20
Romantic	200	200	200	25.48 hrs	912,845	0.0681	160	20	20
Late Romantic	200	200	200	23.95 hrs	789,612	0.0645	160	20	20
Impressionist	200	200	200	20.73 hrs	641,208	0.0583	160	20	20
Modern	200	200	200	22.18 hrs	733,108	0.0710	160	20	20
TOTALS	1,200	1,200	1,200	133.30 hrs	4,472,312	0.0642	960	120	120

---

🛠️ The 3-Stage Acoustic & Symbolic Quality Engine

Every single track in SONATA passed through an exhaustive, multi-gate physics and mastering pipeline before inclusion:

[ Raw MIDI Stream ]
        │
        ▼
[ Gate 1: SMF Parser & Channel 10 Kill-Switch ] ──► (Drops percussion, sysex, running status errors)
        │
        ▼
[ Gate 2: "Clever Hans" Timbral Impostor Purge ] ──► (Purges guitar [Prog 24], strings, flute, choir)
        │
        ▼
[ Gate 3: SoundFont 2.04 Velocity Cross-Fading ] ──► (4-zone multisampling, resonant biquad filtering)
        │
        ▼
[ Gate 4: Damper Tail Decay Modeling (+0.5s) ] ──► (A(t) = A₀ e^(-γt) cos(ωt); avoids Gibbs step artifacts)
        │
        ▼
[ Gate 5: Polyphase Resampling & Mastering ] ────► (Standardized 44.1 kHz, 16-bit PCM, 0.95 peak ceiling)
        │
        ▼
[ Gate 6: Deep Verification & Synchronization ] ──► (0 missing, 0 silent, 0 dupes, |Δt| < 3.5s)

Why Standard Augmentation Fails (And Why SONATA Solves It)

• Pitch Shifting: Shifting an Impressionist piece (e.g., Debussy's Voiles, built on a C-whole-tone hexatonic collection) by $\pm 2$ semitones preserves the interval vector $[0,2,4,6,8,10]$, but completely destroys canonical key-color associations and fundamental formant registers.
• Time Stretching: Compressing or dilating tempo alters note density and pulse rate, but cannot synthesize contrapuntal syntax, voice-leading rules, or cadence morphology.
• SONATA's Solution: True, compositional supplemental ingestion of authentic historical scores, rendered through physical grand piano acoustics.

---

📂 Repository & Dataset Layout

SONATA (Augmented MAESTRO)/
├── audio/
│   ├── baroque/                 # 200 .wav files (44.1 kHz, 16-bit stereo PCM)
│   ├── classical/               # 200 .wav files
│   ├── romantic/                # 200 .wav files
│   ├── late_romantic/           # 200 .wav files
│   ├── impressionist/           # 200 .wav files
│   └── modern/                  # 200 .wav files
├── midi/
│   ├── baroque/                 # 200 .midi files (General MIDI Type 0/1)
│   ├── classical/               # 200 .midi files
│   ├── romantic/                # 200 .midi files
│   ├── late_romantic/           # 200 .midi files
│   ├── impressionist/           # 200 .midi files
│   └── modern/                  # 200 .midi files
├── sonata_dataset.csv           # Master Manifest (MAESTRO v3.0.0 compatible schema)
├── SONATA_Dataset_Report.pdf    # 5-Page Formal Publication Report
└── README.md

Master Manifest Schema (sonata_dataset.csv)

Column	Type	Description
canonical_composer	string	Normalized composer name (e.g., Claude Debussy, Domenico Scarlatti)
canonical_title	string	Official movement / work catalog designation
split	string	Partition designation: train (80%), validation (10%), test (10%)
year	int	Performance capture or catalog year
midi_filename	string	Relative path to symbolic MIDI file (midi/<era>/<file>.midi)
audio_filename	string	Relative path to master acoustic WAV (audio/<era>/<file>.wav)
duration	float	Exact audio playback duration in seconds
era	string	Class label (baroque, classical, romantic, late_romantic, impressionist, modern)

---

⚡ Quickstart: Loading SONATA with PyTorch

After downloading the archive and extracting it, you should have a folder structure like this:

SONATA/
├── audio/
├── midi/
├── balanced_dataset.csv
├── sonata_dataset.csv
├── README.md
└── SONATA_Dataset_Report.pdf

The key point is that the dataset is meant to be used from the extracted root directory named SONATA. Do not assume the files are placed directly in your current working directory.

1. Dual-Stream Multi-Modal PyTorch Dataset

import torch
from torch.utils.data import Dataset, DataLoader
import torchaudio
import pretty_midi
import pandas as pd
from pathlib import Path

class SonataDataset(Dataset):
    def __init__(self, root_dir, split="train", sample_rate=44100, n_mels=128):
        self.root_dir = Path(root_dir)
        df = pd.read_csv(self.root_dir / "sonata_dataset.csv")
        self.df = df[df["split"] == split].reset_index(drop=True)
        self.sample_rate = sample_rate

        self.mel_spectrogram = torchaudio.transforms.MelSpectrogram(
            sample_rate=sample_rate,
            n_fft=2048,
            hop_length=512,
            n_mels=n_mels,
        )

        self.era_to_idx = {
            "baroque": 0,
            "classical": 1,
            "romantic": 2,
            "late_romantic": 3,
            "impressionist": 4,
            "modern": 5,
        }

    def __len__(self):
        return len(self.df)

    def __getitem__(self, idx):
        row = self.df.iloc[idx]

        wav_path = self.root_dir / row["audio_filename"]
        waveform, _ = torchaudio.load(wav_path)
        if waveform.shape[0] > 1:
            waveform = torch.mean(waveform, dim=0, keepdim=True)
        mel_spec = torch.log(self.mel_spectrogram(waveform) + 1e-6)

        midi_path = self.root_dir / row["midi_filename"]
        pm = pretty_midi.PrettyMIDI(str(midi_path))
        piano_roll = torch.from_numpy(pm.get_piano_roll(fs=50)).float()

        label = torch.tensor(self.era_to_idx[row["era"]], dtype=torch.long)

        return {
            "mel_spectrogram": mel_spec,
            "piano_roll": piano_roll,
            "label": label,
            "composer": row["canonical_composer"],
            "title": row["canonical_title"],
        }

# Example usage after extracting SONATA.zip
root_dir = Path("SONATA")
dataset = SonataDataset(root_dir=root_dir, split="train")
loader = DataLoader(dataset, batch_size=16, shuffle=True)
print(f"Loaded {len(dataset)} training tracks across 6 balanced eras.")

If the archive is extracted somewhere else, just replace Path("SONATA") with the full extracted folder path, such as Path("C:/Users/you/Downloads/SONATA").

2. Minimal CSV loading check

import pandas as pd
from pathlib import Path

root_dir = Path("SONATA")
df = pd.read_csv(root_dir / "sonata_dataset.csv")
print(df.head())
print(df["era"].value_counts().sort_index())

This ensures the extracted zip is valid and that the manifest file points to real audio/... and midi/... paths that exist inside the same folder.

---

📊 Scientific Verification & Benchmark Audit

The deep automated auditor evaluated all 1,200 tracks and 2,400 physical files against strict publication criteria:

Verification Metric	Target Threshold	Measured Result	Audit Status
Physical File Completeness	0 missing files	0 missing (2,400 / 2,400 files present)	PASSED ✅
Audio Format Compliance	44.1 kHz, 16-bit stereo PCM	1,200 / 1,200 exact match	PASSED ✅
Silence Detection	Peak $\ge 0.05$, RMS $\ge 0.003$	0 silent tracks (Min peak: 0.125, Mean: 0.863)	PASSED ✅
Audio-MIDI Duration Sync	$	\Delta t	< 3.5\text{ s}$
Duplicate Paths / Hashes	0 duplicates	0 duplicates found	PASSED ✅
Class Distribution Skew	Exactly 200 per class	200 / 200 / 200 / 200 / 200 / 200	PASSED ✅
Split Symmetry	Exactly 160 / 20 / 20 per era	160 train / 20 val / 20 test per era	PASSED ✅

---

📜 Citation & Attribution

If you use SONATA (Augmented MAESTRO) in your research, thesis, or commercial models, please cite:

@dataset{sonata_augmented_maestro_2026,
  title     = {SONATA (Augmented MAESTRO): A Hybrid CNN–Transformer–VAE Framework for Classical Music Era Classification},
  author    = {Aryan Rahman Mugdha and Saadman Bin Jashim},
  year      = {2026},
  publisher = {GitHub},
  note      = {A balanced multimodal classical music dataset combining MAESTRO, ASAP, ADL Piano MIDI, and the Classical MIDI Archive}
}

Underlying Datasets & Foundational Citations

• MAESTRO: Hawthorne et al., "Enabling Factorized Piano Music Modeling and Generation with the MAESTRO Dataset", ICLR 2019.
• ASAP Dataset: Foscarin et al., "ASAP: a dataset of aligned scores and piano performances", ISMIR 2020.
• ADL Piano Dataset: Ferreira et al., "Learning to Generate Music with Sentiment", ISMIR 2019.
• Class Imbalance in Deep Learning: Johnson & Khoshgoftaar, "Survey on deep learning with class imbalance", Journal of Big Data 2019.

---

Built with scientific rigor for the next generation of Music AI.