SECD: String Ensemble Chords Dataset

SECD is a large-scale controlled compositional audio dataset for research on string-ensemble harmony and performance attributes. It contains 287,088 harmonic-interval and chord instances constructed through additive superposition of professionally recorded isolated string notes from the Philharmonia Orchestra into duo-, trio-, and quartet-like four-voice mixtures.

The dataset is organised into six configurations covering harmonic intervals, triads, and seventh chords under loose and strict metadata-consistency conditions. Each mixture includes detailed per-voice metadata, including absolute pitch, dynamic marking, playing technique, source filenames, and audible-duration information.

SECD is designed for Music Information Retrieval (MIR), audio classification, representation learning, chord recognition, playing-technique analysis, dynamics classification, and related research on harmonic, timbral, dynamic, and textural attributes in classical string audio.

Each audio mixture is approximately 2 seconds, mono, and sampled at 22.05 kHz.

Dataset configurations

The Hugging Face release contains six configurations:

Configuration	Examples
harmonic_intervals_loose	90,000
harmonic_intervals_strict	11,224
triads_loose	80,000
triads_strict	4,586
7th_chords_loose	92,598
7th_chords_strict	8,680
Total	287,088

Each configuration contains a single split named train.

The name train here means all examples belonging to that configuration. SECD does not define one universal intrinsic train/validation/test partition. Evaluation partitions depend on the research task and experimental protocol.

Strict and loose subsets

SECD distinguishes between strict and loose subsets.

In the strict subsets, the instruments participating in a mixture share the same dynamic and playing-technique condition.

In the loose subsets, participating instruments may have different dynamics and/or playing techniques. The available per-instrument metadata should therefore be used when studying performance attributes.

For harmonic intervals, the original source metadata contains shared textual dynamic and technique fields together with instrument-specific encoded attributes. In loose interval subsets, the shared textual fields should not be interpreted as proof that both instruments necessarily use identical performance attributes.

Dataset structure

Harmonic intervals

The harmonic-interval configurations contain duets formed from the following instrument pairs:

• cello–viola
• cello–violin 1
• viola–violin 1

Metadata includes pitch labels, interval names, semitone distance, original source filenames, instrument codes, pitch codes, technique codes, dynamic codes, audible durations, and the mixture filename.

Because the three instrument-pair CSV schemas differ, the Hugging Face representation uses a union schema. Fields that do not apply to a particular instrument pair are null.

Triads

Triads contain cello, viola, and violin 1.

The dataset includes four chord qualities:

• major
• minor
• diminished
• augmented

Metadata includes chord name, theoretical spelling, root pitch, chord quality, inversion, open/closed position, per-instrument pitches, source filenames, dynamics, techniques, audible durations, and the mixture filename.

Seventh chords

Seventh chords contain cello, viola, violin 2, and violin 1.

The source quality labels are preserved exactly in the Hugging Face representation, including:

• 7 — dominant seventh
• maj7 — major seventh
• min7 — minor seventh
• m7b5 — half-diminished seventh

Metadata includes theoretical spelling, chord name, root pitch, quality, inversion, position, per-instrument pitches, source filenames, dynamics, techniques, audible durations, and the mixture filename.

Schema note

The harmonic-interval configurations require a unified schema because the three source instrument-pair CSVs contain different instrument-specific columns. Consequently, fields corresponding to instruments not present in a given duet are stored as null.

This affects only the representation of pair-specific fields and does not limit filtering or analysis by interval, semitone distance, or other shared metadata across the three duet types.

Audio representation

The Hugging Face version stores mixture audio directly in Parquet using the Hugging Face Audio feature:

Audio(sampling_rate=22050)

This allows direct ML-oriented loading without separately resolving hundreds of thousands of individual WAV files.

The original source filenames are retained as provenance metadata.

Loading SECD

Install the Hugging Face datasets library:

pip install datasets

Load one configuration:

from datasets import load_dataset

ds = load_dataset(
    "ageroul/SECD",
    "triads_strict"
)

print(ds)
print(ds["train"].features)

Load another configuration:

ds = load_dataset(
    "ageroul/SECD",
    "harmonic_intervals_loose"
)

Inspect an example without forcing local audio decoding:

from datasets import Audio

ds_no_decode = ds.cast_column(
    "audio",
    Audio(sampling_rate=22050, decode=False)
)

example = ds_no_decode["train"][0]

print(example["audio"])

To decode the audio waveform, use the default Audio feature. This may require compatible audio-decoding dependencies in the local environment.

Hugging Face release and archival release

This Hugging Face repository is the ML-native distribution layer of SECD.

The canonical archival and versioned maintenance release remains available through Zenodo:

Dataset DOI:
https://doi.org/10.5281/zenodo.22728050

The maintenance release corrects octave numbering for enharmonic B#/Cb spellings in theory_name across 12 metadata CSVs. The other 10 metadata CSVs are unchanged. Audio content, WAV filenames and paths, schemas, row counts, and all other metadata fields are unchanged.

Zenodo should be used when an archival citation, original release package, or preserved source distribution is required.

The Hugging Face version is intended to make SECD easier to discover, inspect, load, and integrate into machine-learning workflows.

Associated publication

SECD is described in the peer-reviewed article:

Geroulanos, A.; Zervas, P.; Tzimas, G. SECD: A String Ensemble Chords Dataset for Multi-Task Audio Classification. Acoustics 2026, 8, 48. https://doi.org/10.3390/acoustics8030048

Researchers using SECD are encouraged to cite both the dataset release and the associated publication where appropriate.

License

SECD is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Users may share and adapt the material provided that appropriate attribution is given according to the license terms.

Reproducibility and provenance

The Hugging Face conversion preserves the original source metadata and source filenames while adding:

• embedded mixture audio using the Hugging Face Audio feature;
• a source_csv field identifying the originating metadata file.

The conversion was performed deterministically from the canonical SECD maintenance release.

The six Hugging Face configurations contain exactly:

90,000   harmonic_intervals_loose
11,224   harmonic_intervals_strict
80,000   triads_loose
4,586    triads_strict
92,598   7th_chords_loose
8,680    7th_chords_strict
-------
287,088  total

Notes for benchmark design

SECD supports multiple classification and representation-learning tasks. Researchers should define splits appropriate to the target task.

In particular, evaluation protocols that aim to measure generalization to unseen source recordings should consider source-note-disjoint partitioning rather than relying only on mixture-level separation.

The Hugging Face train split should therefore be interpreted as the complete configuration pool, not as a prescribed experimental training partition.

Resources

• Dataset archive: https://doi.org/10.5281/zenodo.22728050
• Dataset concept DOI: https://doi.org/10.5281/zenodo.15547206
• Associated article: https://doi.org/10.3390/acoustics8030048
• Source repository: https://github.com/ageroul/SECD