asigalov61/Mono-Segments

Mono Segments

SOTA MIDI dataset which contains 310k+ select multi-instrumental MIDIs from Discover MIDI Dataset with lead monophonic melodies and music segments labels

---

Abstract

Mono Segments is a large‑scale, state‑of‑the‑art symbolic music dataset comprising 310k+ carefully selected multi‑instrumental MIDI files enriched with lead monophonic melodies and high‑precision structural segment labels. Built from the Discover MIDI Dataset, it provides a uniquely comprehensive and genre‑diverse corpus designed specifically for advancing music‑segmentation AI, symbolic analysis, and sequence‑modeling research. Each MIDI includes standardized text_event annotations—Count‑In, Intro, Segment #, Bridge #, Outro—encoded in absolute milliseconds and normalized to begin at time zero, ensuring consistent alignment across tempo variations and seamless compatibility with MIDI processors and karaoke‑style playback systems.

The dataset preserves full instrumental arrangements, including percussion, enabling models to learn segmentation cues from rich musical textures while also supporting optional derivation of solo‑piano variants to expand training sets. Segment labels are generated through a refined algorithmic pipeline achieving approximately 99% accuracy, with fewer than 1% of files expected to contain malformed structures or imperfect annotations. To maintain traceability and reproducibility, each MIDI retains the original MD5 hash from the source dataset, allowing retrieval of the corresponding unnormalized file when needed.

By combining scale, structural clarity, stylistic breadth, and practical labeling conventions, Mono Segments offers a robust foundation for developing next‑generation symbolic music segmentation systems, facilitating both high‑level structural modeling and fine‑grained temporal analysis in research and production environments.

---

Spotlight Features

• First of its kind large-scale high-qualiy segmented MIDI dataset
• Almost perfect segments labels (99% accurate)
• Diverse and comprehensive music corpus that covers all concievable music genres and styles
• Multi-instrumental MIDIs with drums for utlimate utility and versatility

---

Intended use

• The dataset was designed primarily for creating symbolic music (MIDI) segmentation AI models

---

Labels info

• Segments labels are included in each MIDI in a form of standard MIDI text_event which can be easily read with any MIDI processor or a Karaoke MIDI player
• Labels are named to indicate a corresponding type of a music segment: Count-In, Intro, Segment #, Bridge #, Outro
• Count-in, Intro and Outro labels may not be present if corresponding segments do not exist
• All MIDIs and all labels are timed in absolute milliseconds and shifted to start with 0

---

Pro Tips

• If you are training multi-instrumental AI model you can easily double the size of the dataset by creating a solo Piano version of the full dataset
• If you want to use segments labels from the dataset with the original MIDIs, make sure to convert original MIDIs to milliseconds (including tempo changes, etc) and shift all events to start with 0 or the labels will not align

---

Limitations

• Dataset contains normalized barebone/basic versions of the original MIDIs. However, each MIDI includes the original MD5 hash (MIDI name) from Discover MIDI dataset so that original MIDIs can be retrieved and used if needed.
• Music segments labels may not be 100% accurate due to algorithmic labeling, so make sure to double-check the labels prior to use
• Some MIDIs may be malformed so please double check the structure prior to use
• In both cases there should be less than 1% of malformed MIDIs/incorrect labels

---

Citation

@misc{Lev2026MonoSegments,
  author       = {Alex Lev},
  title        = {Mono Segments MIDI Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/asigalov61/Mono-Segments}},
  note         = {Project Los Angeles, Tegridy Code}
}

---

Project Los Angeles

Tegridy Code 2026