OpenWhistleNeurIPS26/OpenWhistle-Detection-Finetuning

OpenWhistle Detection Finetuning

Expert-annotated whistle-type detection dataset for OpenWhistle. Each example is a fixed-length 0.5 s audio window labeled with the whistle types present in that window; background/no-whistle windows are represented by an all-zero target vector.

Overview

• Task: multi-label whistle-type detection on fixed-length audio windows
• Target vector: label, with one binary decision per whistle type in the order SW_Neo, SW_Luna, SW_Nikita, SW_Nana, SW_Yosefa, SW_Dana, NSW_1
• Background: no-whistle windows use an all-zero label vector
• Convenience binary label: binary_label, derived from label, with values noise and whistle
• Total rows: 5600
• Class balance: 400 examples per whistle type across 7 classes, plus 2800 background windows
• Split: session-disjoint train/validation/test with ratios 0.70/0.15/0.15

Construction

• Built from the expert-annotated OpenWhistle benchmark subset
• Multi-label target: encode each whistle type present in the window in the fixed-length label vector
• Binary target: derive whistle when at least one label dimension is active, otherwise noise
• Session key: source original_path basename with audio extension removed
• Split rule: exact MILP assignment of whole sessions
• Balance rule: exact source-label targets per split

Features

• audio: audio clips stored with decode=False
• label: multi-label whistle-type vector in order SW_Neo, SW_Luna, SW_Nikita, SW_Nana, SW_Yosefa, SW_Dana, NSW_1
• binary_label: derived binary class label with values noise and whistle
• name: original clip filename from the source dataset
• source_label: original source label before deriving the binary label

Rows By Split

• train: 3920 rows
• validation: 840 rows
• test: 840 rows

Binary Label Counts

• total: noise=2800, whistle=2800
• train: noise=1960, whistle=1960
• validation: noise=420, whistle=420
• test: noise=420, whistle=420

Source Label Counts

• train: Dana=280, Luna=280, NSW_1=280, Nana=280, Neo=280, Nikita=280, Yosefa=280, noise=1960
• validation: Dana=60, Luna=60, NSW_1=60, Nana=60, Neo=60, Nikita=60, Yosefa=60, noise=420
• test: Dana=60, Luna=60, NSW_1=60, Nana=60, Neo=60, Nikita=60, Yosefa=60, noise=420

Session Leakage

• Pairwise session overlap: train__validation=0, train__test=0, validation__test=0

Example

from datasets import Audio, load_dataset

dataset = load_dataset("OpenWhistleNeurIPS26/OpenWhistle-Detection-Finetuning")
decoded_train = dataset["train"].cast_column("audio", Audio())
sample = decoded_train[0]