projectlosangeles/Annotated-MIDI-Songs

Annotated MIDI Songs

High‑Resolution SongFormer Structural Annotations with MIDI/MP3 Pairs

---

Overview

The dataset contains 26k+ MIDI songs with SongFormer‑generated structural segment annotations, paired with high‑quality MP3 renderings for each MIDI file. The collection spans all major music genres and styles, providing a unified, richly annotated symbolic+audio resource for Music AI, MIR, sequence modeling, and structure‑aware generative systems.

---

🌟 Key Features

🎼 SongFormer Structural Annotations

Each song includes a JSON file containing SongFormer model–generated segment boundaries. SongFormer (Hao et al., 2026) introduced a scalable transformer architecture trained with heterogeneous supervision, enabling high‑quality music structure detection across diverse genres.

The dataset adopts the same annotation taxonomy:

• silence
• intro
• verse
• chorus
• bridge (when detected)
• outro

Each segment includes precise floating‑point timestamps (start, end) aligned to the MP3 audio and MIDI performance.

🎧 MIDI + MP3 Pairs

Every MIDI file has a corresponding MP3 rendering:

• Enables audio‑symbolic alignment research
• Supports audio‑to‑MIDI, MIDI‑to‑audio, and cross‑modal learning
• Provides a unified dataset for SongFormer fine‑tuning, audio transformers, and symbolic‑audio joint models

🎹 Oroginal MIDI Files

Dataset contains pure original MIDIs without any modifications

🧩 High‑Precision Segment JSONs

Example annotation:

{
  "label": "chorus",
  "start": 23.400936037441497,
  "end": 42.241689667586705
}

Segments are:

• non‑overlapping
• chronologically ordered
• SongFormer‑taxonomy compliant
• aligned to both MIDI and MP3 timelines

---

📁 Dataset Structure

├── Annotations/   # SongFormer annotations
├── Code/          # Tools for MIDI parsing, rendering, and annotation processing
├── Metadata/      # Detailed metadata for each MIDI
├── MIDIs/         # Original MIDI files
├── MP3s/          # High-quality audio renderings aligned to each MIDI
└── SoundFont/     # HQ Sound Font 2 bank that was used to render all MIDIs

For each song:

MIDI_MD5_Hash.mid    # Original MIDI
MIDI_MD5_Hash.mp3    # HQ MIDI rendering as 320 kbps MP3
MIDI_MD5_Hash.json   # SongFormer structural annotations

---

🔍 Annotation Format

Each annotation file is a list of segments:

[
  {"label": "silence", "start": 0.0, "end": 4.560182407296292},
  {"label": "intro",   "start": 4.560182407296292, "end": 23.400936037441497},
  {"label": "chorus",  "start": 23.400936037441497, "end": 42.241689667586705},
  {"label": "verse",   "start": 42.241689667586705, "end": 61.082443297731906},
  ...
]

Annotation Guarantees

• Temporal precision suitable for alignment tasks
• Uniform label taxonomy across all 26k songs
• Direct compatibility with SongFormer training pipelines
• Supports long‑context transformer models
• Ideal for MIR benchmarking and structural segmentation research

---

🧠 SongFormer Annotation Details

SongFormer (Hao et al., 2026) is a transformer‑based architecture trained with heterogeneous supervision, combining:

• symbolic cues
• audio alignment signals
• structural priors
• large‑scale weak labels

The dataset was created using official SongFormer model to generate:

• section boundaries
• section types
• silence detection
• multi‑segment hierarchical structure (when present)

This makes the dataset uniquely suited for:

• structure‑aware music generation
• form‑conditioned transformers
• audio‑symbolic multimodal models
• automatic form detection
• music segmentation benchmarking

---

📚 Usage Examples

Load MIDI + Annotation

import json
from mido import MidiFile

midi = MidiFile("MIDIs/MIDI_MD5_Hash.mid")
with open("Annotations/MIDI_MD5_Hash.json") as f:
    segments = json.load(f)

for seg in segments:
    print(seg["label"], seg["start"], seg["end"])

Align MIDI + MP3

import librosa

audio, sr = librosa.load("MP3s/MIDI_MD5_Hash.mp3")
# Use segment timestamps to slice audio or align symbolic events

---

📖 Citations

Please cite the following when using this dataset:

@misc{Lev2026TegridyMIDIDataset,
  author       = {Alex Lev},
  title        = {Tegridy MIDI Dataset: A Comprehensive Multi-Instrumental MIDI Resource for Music AI and MIR Research},
  year         = {2026},
  howpublished = {\url{https://github.com/asigalov61/Tegridy-MIDI-Dataset}},
  organization = {GitHub},
  note         = {Includes original Tegridy collections, curated links to major external datasets (Discover, Godzilla, MAESTRO, Lakh, etc.), software tools, and AI applications. Licensed under CC BY-NC-SA 4.0.}
}

@phdthesis{raffel2016learning,
  author       = {Colin Raffel},
  title        = {Learning-Based Methods for Comparing Sequences, with Applications to Audio-to-{MIDI} Alignment and Matching},
  school       = {Columbia University},
  year         = {2016},
  url          = {https://colinraffel.com/projects/lmd/}
}

@misc{hao2026songformer,
  title         = {SongFormer: Scaling Music Structure Analysis with Heterogeneous Supervision},
  author        = {Hao, Chunbo and Yuan, Ruibin and Yao, Jixun and Deng, Qixin and Bai, Xinyi and Wang, Yanbo and Xue, Wei and Xie, Lei},
  year          = {2026},
  eprint        = {2510.02797},
  archivePrefix = {arXiv},
  primaryClass  = {eess.AS},
  url           = {https://arxiv.org/abs/2510.02797}
}

---

📜 License

CC BY‑NC‑SA 4.0
Non‑commercial use permitted with attribution. Derivatives must share the same license.

---

Project Los Angeles

Tegridy Code 2026