---
pretty_name: ALIGN Dataset
task_categories:
  - audio-classification
tags:
  - music
  - clarinet
  - musicxml
  - music-performance
  - error-detection
  - synthetic
  - alignment
size_categories:
  - 10K<n<100K
---

# ALIGN Dataset

Synthetic **Bb clarinet** practice bundles for ALIGN / MusicEval: a clean MusicXML score paired with a SoundFont performance that contains planted faults. Gold labels are **score-part melodies** (schema **1.2**), not only wall-clock intervals.

The clean score stays correct. Only the synthesized performance is wrong.

## Splits

Samples are packed into zip shards (250 clips each) so the Hub stays within file-count limits.

| Folder | Samples | Source scores |
|--------|--------:|---------------|
| `output/part_XXXX.zip` | 12,000 | Procedural original clarinet scores |
| `output_2k_rawdata/part_XXXX.zip` | 2,104 | Random measure windows from uploaded `RawData/Score` MusicXML |

`manifest.json` lists which sample ids sit in each shard.

After unzipping a shard you get one directory per clip, for example `synth_gen_0042/`:

| File | Role |
|------|------|
| `verified_score.musicxml` | Clean notation (ground truth) |
| `performance_score.musicxml` | Errored (and possibly repeated) render source |
| `reference_audio.wav` | Clarinet render of the clean score |
| `performance_audio.wav` | Clarinet render with planted errors |
| `reference_audio.mid` / `performance_audio.mid` | MIDI sent to the SoundFont |
| `labels.json` | Synthetic gold (`source: synthetic`, schema 1.2) |
| `candidates.json` | Auto alignment candidates |
| `alignment.npz` | DTW alignment artifacts |
| `performance_mel.npy` / `reference_mel.npy` | Log-mel features (22.05 kHz, 128 mels) |
| `metadata.json` | Render settings, error types, seed |

Audio is **sounding pitch** (`sounding_transpose: -2`): written C sounds Bb. MusicXML and `labels.json` `pitches` stay **written**.

## Planted errors

A clip draws **1–8** non-overlapping content errors from five equally weighted types, then often repeats the affected measure(s):

| Type | What changes in the performance |
|------|----------------------------------|
| `wrong_note` | ±1 or ±2 semitones, or a squeak (MIDI C6–A7) |
| `missed_note` | Written note replaced by a rest |
| `extra_note` | Split a note; insert a neighbor or squeak |
| `intonation_error` | Pitch bend 40–80 cents on 1–4 notes |
| `rhythm_error` | Late/early start or end, tempo change, or uneven durations |

**Repetition** is applied after content errors (usual case) or as a standalone replay of a sounding measure. A silent 0.2–1.0 s gap is inserted before the copy. Repeated-pass copies of content labels are **not** gold.

Melody gold is a contiguous run of clean-score notes (`score_part` + `pitches`) with 1–2 notes of pad on each side.

## Loading

```python
from huggingface_hub import hf_hub_download
from zipfile import ZipFile
from pathlib import Path
import json

zip_path = hf_hub_download(
    repo_id="LapisLa2uli/ALIGN_Dataset",
    filename="output/part_0000.zip",
    repo_type="dataset",
)
out = Path("align_samples")
with ZipFile(zip_path) as zf:
    zf.extractall(out)
sample = next(p for p in out.iterdir() if p.is_dir())
labels = json.loads((sample / "labels.json").read_text(encoding="utf-8"))
print(sample.name, len(labels["labels"]))
```

## Notes

- Render: tinysoundfont + FreePats clarinet SoundFont; MIDI via music21.
- Official synth eval in ALIGN is contiguous pitch-list **containment F1**, not timestamp IoU.
- `pipeline.log` files are omitted from the upload (they contain local machine paths).