REMI-z 0.7.4

REMI‑z: Tokenizer & MultiTrack Music Data Structure

REMI‑z is an official implementation of the tokenizer proposed in
Unifying Symbolic Music Arrangement: Track-Aware Reconstruction and Structured Tokenization.

Paper · Demo · GitHub · PyPI · Author

REMI‑z provides:

• Efficient sequence representation of multitrack symbolic music.
• Bar‑level control over structure, instrumentation and content.
• Round‑trip conversion between MIDI ⇄ REMI‑z tokens, plus piano‑roll and chord utilities.

Core Concepts

At the heart of REMI‑z is a hierarchical data structure:

• MultiTrack – the whole piece, a list of bars.
  • Bar – one bar of music with its time signature and tempo, holding multiple tracks.
    • Track – all notes of one instrument in that bar (e.g. piano, strings, drums).
      • Note – a single note (onset, duration, pitch, velocity).

You can freely convert this hierarchy to and from:

• MIDI files (via MultiTrack.from_midi(...) / MultiTrack.to_midi(...))
• REMI‑z token sequences (via MultiTrack.from_remiz_str(...) / MultiTrack.to_remiz_str(...))
• Piano‑roll matrices and content sequences at bar or song level.

Installation

From PyPI

pip install REMI-z

From source

git clone https://github.com/Sonata165/REMI-z.git
cd REMI-z
pip install -r Requirements.txt
pip install -e .

Quickstart

1. MIDI → REMI‑z tokens

from remi_z import MultiTrack

mt = MultiTrack.from_midi("your_song.mid")
remiz_str = mt.to_remiz_str(with_ts=True, with_tempo=True, with_velocity=True)

2. REMI‑z tokens → MIDI

from remi_z import MultiTrack

mt = MultiTrack.from_remiz_str("your_remiz_string")
mt.to_midi(remiz_str, "reconstructed.mid")

3. Bar‑level manipulation

from remi_z.multitrack import MultiTrack

mt = MultiTrack.from_midi("your_song.mid")

spans_of_bar_1to3 = mt[1:4] # MultiTrack object
bar_10 = mt[10] # Bar object

3. Other useful manipulation

from remi_z.multitrack import MultiTrack

mt = MultiTrack.from_midi("your_song.mid")

# Normalize key to C major or A minor
mt.key_norm()

# Transpose all non‑drum tracks up a whole tone
mt.shift_pitch(2)

# Quantize to 16th notes
mt.quantize_to_16th()

# Set velocity to specific track
mt.set_velocity(velocity=64, track_id=0)

# Set tempo
mt.set_tempo(90)

# Detect chord
bar = mt[0]
bar.get_chord()

4. Performance MIDI — NoteStream & MultiStream

For performance MIDI (absolute timing in seconds; no bar/time‑signature/tempo structure) use NoteStream (single track) and MultiStream (multi‑track).

from remi_z import NoteStream, MultiStream

# ── Single track ──────────────────────────────────────────────────────────────
ns = NoteStream.from_midi("single_track.mid")            # one instrument track
ns = NoteStream.from_triplet_list([[0.0, 0.5, 60],       # or from [onset, offset, pitch]
                                   [0.5, 1.0, 62]])
triplets = ns.to_triplet_list()                          # -> [[onset, offset, pitch], ...]
ns.to_midi("single_out.mid", program=0)

# ── Multi-track ───────────────────────────────────────────────────────────────
ms = MultiStream.from_midi("multitrack.mid", skip_drums=False, dedup=False)
print(len(ms), ms.programs)        # number of tracks + their GM programs (128 = drums)
ms.to_midi("multitrack_out.mid")   # one instrument per track, programs restored

# Build from individual tracks (several tracks may share the same program)
ms = MultiStream.from_note_streams([ns, NoteStream.from_midi("other.mid")])
groups = ms.by_program()           # {program: [NoteStream, ...]} (duplicate-safe)

# Flatten all pitched tracks into one deduplicated single-track stream
flat = ms.flatten()                # drops drums; keeps the longest note per (onset, pitch)
flat.to_midi("flattened.mid")

Tip: reading a multi-instrument MIDI directly into one NoteStream requires NoteStream.from_midi(path, merge_tracks=True); otherwise use MultiStream.

For more examples, see demo.ipynb.

Feature Highlights

• Tokenizer

  • REMI‑z representation with rich bar structure and instrument awareness.
  • Supports time‑signature and tempo tokens (s-*, t-*) and optional velocity tokens (v-*).

• Hierarchy & Manipulation

  • Note / Track / Bar / MultiTrack classes for structured symbolic music editing.
  • Bar‑level flattening, track filtering, instrument remapping, phrase permutation.
  • Melody extraction and global pitch‑range analysis.

• I/O & Interop

  • Robust MIDI import/export (supports tempo and time‑signature changes).
  • Support for multiple tracks with the same program ID (e.g., multiple piano parts).
  • Piano‑roll conversion and content sequences for modeling tasks.

• Performance MIDI (absolute timing)

  • NoteStream — a single instrument track of absolute‑time notes (seconds), with from_midi / from_triplet_list / to_midi / to_triplet_list; carries its program in inst_id (128 = drums).
  • MultiStream — an ordered collection of NoteStream tracks (handles multiple tracks of the same program), with from_midi / from_note_streams / to_midi, by_program() grouping, and flatten() to a single deduplicated track.
  • Ideal for transcription ground truth and onset/offset/pitch workflows where bar/tempo structure isn't needed.

• Harmony Tools

  • Bar‑wise half‑bar chord detection (Bar.get_chord()).
  • Chord objects and chord sequences (Chord, ChordSeq) with REMI‑z pitch token generation.

Updates

• 2026‑06‑29 – Add performance‑MIDI support: NoteStream (single‑track, absolute‑time notes) and MultiStream (multi‑track), with MIDI/triplet I/O, drum skipping, duplicate‑note removal, and flattening to a single track.
• 2025‑09‑09 – Support reading and writing MIDIs containing multiple tracks of the same program ID.

Known Issues

• Creating a MultiTrack object from REMI‑z sequence does not yet support multiple tracks of the same program ID.

Citation

If you use REMI‑z in your research, please cite:

@inproceedings{ou2025unifying,
    title     = {Unifying Symbolic Music Arrangement: Track-Aware Reconstruction and Structured Tokenization},
    author    = {Ou, Longshen and Zhao, Jingwei and Wang, Ziyu and Xia, Gus and Liang, Qihao and Hopkins, Torin and Wang, Ye},
    booktitle = {Proceedings of the 39th Conference on Neural Information Processing Systems (NeurIPS)},
    year      = {2025}
}

致中文用户（简要说明）

REMI‑z 是一个面向符号音乐编曲/生成的多轨分词与数据结构库：

• 支持在 小节层面 操作 MIDI（如乐器筛选、移调、量化、合并/拆分轨道等）。
• 提供 MIDI ⇄ REMI‑z token 的双向转换，以及钢琴卷帘、和弦/旋律提取等工具。
• 推荐先安装后查看 demo.ipynb 来快速上手。