Skip to content

nexuml_library.data.dcaset2.dcase2026

nexuml_library.data.dcaset2.dcase2026

DCASE 2026 dataset sources for Tasks 1, 2, and 7.

All datasets are metadata-backed and fake-fixture friendly: they accept pre-built DataFrames and tiny local audio files for testing without requiring network downloads.

DCASET2Dataset

Bases: DataSourceDefinition

DCASE Task 2 anomalous sound detection dataset.

Builds a metadata DataFrame from the DCASE directory structure and lazily loads audio/features in load_item().

Directory layout expected

//_data/raw//train/ //_data/raw//test/

machine (int index), y_true (0=normal,1=anomalous),

condition (int), target (0/1).

Attributes:

Name Type Description
root str | Path | None

Path to the dataset root directory.
dataset_name str

e.g. "DCASE2024T2"
machine_type str

e.g. "ToyCar", "fan", etc.
section_keyword str

prefix for section names, e.g. "section"
section_ids list[str] | None

list of section id strings, e.g. ["00", "01"]
data_type str

"dev" or "eval"
train bool

if True load train split, else test split
sample_rate int

target sample rate for audio loading
machine_types list[str] | None

list of all machine types (for machine index mapping)

DCASE2026T1Dataset

Bases: DataSourceDefinition

DCASE 2026 Task 1 dataset for BSD10k-v1.2 / BSD35k-CS.

Expects a metadata CSV with columns including at least: file, sound_id, class_top, class_second. Optional columns: confidence, uploader, license, title, tags, description, clap_embedding.

Attributes:

Name Type Description
root str | Path

Dataset root directory.
metadata_csv str

Path to metadata CSV relative to root.
download_mode DownloadMode

One of the DownloadMode literals.
checksum str | None

Optional expected checksum of the metadata CSV.
clap_embeddings_dir str | None

Optional directory with pre-computed CLAP .npy files.
fold_seed int

Seed for reproducible baseline 5-fold splits.
fold_id int | None

If provided, filter to that fold (0-4).
split str

"fit", "val", "test", or "eval".
sample_rate int

Target audio sample rate.

DCASE2026T2Dataset

Bases: DCASET2Dataset

DCASE 2026 Task 2 dataset — metadata index with domain/attribute enrichment.

Extends DCASET2Dataset with 2026-specific metadata: - domain inferred from filename source/target markers - attributes merged from the official attributes CSV when available

Audio is 2-channel (stereo). Channel selection happens in pipeline layers, not here. DALI loads raw stereo audio as [C, T] via dali_layout="CT".

DCASE2026T7Dataset

Bases: DataSourceDefinition

DCASE 2026 Task 7 DIL-DCASE26 dataset.

Expects a root directory with subfolders D2/ and D3/ (and optionally D1/ for baseline reference). Each domain folder may contain:

  • metadata.csv with columns file, class, domain
  • Raw audio files (optionally with evaluation_setup/ split files)

The dataset exposes a sequence-of-increments interface via get_increment().