ageroul commited on
Commit
1baf239
·
verified ·
1 Parent(s): bb1b165

Complete and validate SECD dataset card

Browse files
Files changed (1) hide show
  1. README.md +128 -3
README.md CHANGED
@@ -104,6 +104,7 @@ dataset_info:
104
  num_examples: 92598
105
  download_size: 8179853294
106
  dataset_size: 8227282409
 
107
  - config_name: 7th_chords_strict
108
  features:
109
  - name: audio
@@ -172,6 +173,7 @@ dataset_info:
172
  num_examples: 8680
173
  download_size: 766782323
174
  dataset_size: 771154860
 
175
  - config_name: harmonic_intervals_loose
176
  features:
177
  - name: audio
@@ -238,6 +240,7 @@ dataset_info:
238
  num_examples: 90000
239
  download_size: 7948054328
240
  dataset_size: 7973170807
 
241
  - config_name: harmonic_intervals_strict
242
  features:
243
  - name: audio
@@ -304,6 +307,7 @@ dataset_info:
304
  num_examples: 11224
305
  download_size: 991188361
306
  dataset_size: 994324900
 
307
  - config_name: triads_loose
308
  features:
309
  - name: audio
@@ -362,6 +366,7 @@ dataset_info:
362
  num_examples: 80000
363
  download_size: 7066768509
364
  dataset_size: 7097369872
 
365
  - config_name: triads_strict
366
  features:
367
  - name: audio
@@ -465,8 +470,8 @@ For harmonic intervals, the original source metadata contains shared textual `dy
465
  The harmonic-interval configurations contain duets formed from the following instrument pairs:
466
 
467
  - cello–viola
468
- - cello–violin
469
- - viola–violin
470
 
471
  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.
472
 
@@ -503,4 +508,124 @@ Metadata includes theoretical spelling, chord name, root pitch, quality, inversi
503
  The Hugging Face version stores mixture audio directly in Parquet using the Hugging Face `Audio` feature:
504
 
505
  ```python
506
- Audio(sampling_rate=22050)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
104
  num_examples: 92598
105
  download_size: 8179853294
106
  dataset_size: 8227282409
107
+
108
  - config_name: 7th_chords_strict
109
  features:
110
  - name: audio
 
173
  num_examples: 8680
174
  download_size: 766782323
175
  dataset_size: 771154860
176
+
177
  - config_name: harmonic_intervals_loose
178
  features:
179
  - name: audio
 
240
  num_examples: 90000
241
  download_size: 7948054328
242
  dataset_size: 7973170807
243
+
244
  - config_name: harmonic_intervals_strict
245
  features:
246
  - name: audio
 
307
  num_examples: 11224
308
  download_size: 991188361
309
  dataset_size: 994324900
310
+
311
  - config_name: triads_loose
312
  features:
313
  - name: audio
 
366
  num_examples: 80000
367
  download_size: 7066768509
368
  dataset_size: 7097369872
369
+
370
  - config_name: triads_strict
371
  features:
372
  - name: audio
 
470
  The harmonic-interval configurations contain duets formed from the following instrument pairs:
471
 
472
  - cello–viola
473
+ - cello–violin 1
474
+ - viola–violin 1
475
 
476
  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.
477
 
 
508
  The Hugging Face version stores mixture audio directly in Parquet using the Hugging Face `Audio` feature:
509
 
510
  ```python
511
+ Audio(sampling_rate=22050)
512
+ ```
513
+
514
+ This allows direct ML-oriented loading without separately resolving hundreds of thousands of individual WAV files.
515
+
516
+ The original source filenames are retained as provenance metadata.
517
+
518
+ ## Loading SECD
519
+
520
+ Install the Hugging Face `datasets` library:
521
+
522
+ ```bash
523
+ pip install datasets
524
+ ```
525
+
526
+ Load one configuration:
527
+
528
+ ```python
529
+ from datasets import load_dataset
530
+
531
+ ds = load_dataset(
532
+ "ageroul/SECD",
533
+ "triads_strict"
534
+ )
535
+
536
+ print(ds)
537
+ print(ds["train"].features)
538
+ ```
539
+
540
+ Load another configuration:
541
+
542
+ ```python
543
+ ds = load_dataset(
544
+ "ageroul/SECD",
545
+ "harmonic_intervals_loose"
546
+ )
547
+ ```
548
+
549
+ Inspect an example without forcing local audio decoding:
550
+
551
+ ```python
552
+ from datasets import Audio
553
+
554
+ ds_no_decode = ds.cast_column(
555
+ "audio",
556
+ Audio(sampling_rate=22050, decode=False)
557
+ )
558
+
559
+ example = ds_no_decode["train"][0]
560
+
561
+ print(example["audio"])
562
+ ```
563
+
564
+ To decode the audio waveform, use the default `Audio` feature. This may require compatible audio-decoding dependencies in the local environment.
565
+
566
+ ## Hugging Face release and archival release
567
+
568
+ This Hugging Face repository is the **ML-native distribution layer** of SECD.
569
+
570
+ The canonical archival and versioned release remains available through Zenodo:
571
+
572
+ **Dataset DOI:**
573
+ `10.5281/zenodo.15547207`
574
+
575
+ Zenodo should be used when an archival citation, original release package, or preserved source distribution is required.
576
+
577
+ The Hugging Face version is intended to make SECD easier to discover, inspect, load, and integrate into machine-learning workflows.
578
+
579
+ ## Associated publication
580
+
581
+ SECD is described in the peer-reviewed article:
582
+
583
+ **SECD: A String Ensemble Chords Dataset for Multi-Task Audio Classification**
584
+ *Acoustics*, 2026.
585
+
586
+ **Article DOI:**
587
+ `10.3390/acoustics8030048`
588
+
589
+ Researchers using SECD are encouraged to cite both the dataset release and the associated publication where appropriate.
590
+
591
+ ## License
592
+
593
+ SECD is distributed under the **Creative Commons Attribution 4.0 International (CC BY 4.0)** license.
594
+
595
+ Users may share and adapt the material provided that appropriate attribution is given according to the license terms.
596
+
597
+ ## Reproducibility and provenance
598
+
599
+ The Hugging Face conversion preserves the original source metadata and source filenames while adding:
600
+
601
+ - embedded mixture audio using the Hugging Face `Audio` feature;
602
+ - a `source_csv` field identifying the originating metadata file.
603
+
604
+ The conversion was performed deterministically from the canonical SECD release.
605
+
606
+ The six Hugging Face configurations contain exactly:
607
+
608
+ ```text
609
+ 90,000 harmonic_intervals_loose
610
+ 11,224 harmonic_intervals_strict
611
+ 80,000 triads_loose
612
+ 4,586 triads_strict
613
+ 92,598 7th_chords_loose
614
+ 8,680 7th_chords_strict
615
+ -------
616
+ 287,088 total
617
+ ```
618
+
619
+ ## Notes for benchmark design
620
+
621
+ SECD supports multiple classification and representation-learning tasks. Researchers should define splits appropriate to the target task.
622
+
623
+ 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.
624
+
625
+ The Hugging Face `train` split should therefore be interpreted as the complete configuration pool, not as a prescribed experimental training partition.
626
+
627
+ ## Resources
628
+
629
+ - Dataset archive: https://doi.org/10.5281/zenodo.15547207
630
+ - Associated article: https://doi.org/10.3390/acoustics8030048
631
+ - Source repository: https://github.com/ageroul/SECD