EPIC-SOUNDS

sounds

General information

Label Value Description
Name EPIC-SOUNDS Full dataset name
ID sounds/epic_sounds Datalist id for external indexing
Abbreviation EPIC-SOUNDS Official dataset abbreviation, e.g. one used in the original paper
Provider University of Oxford, University of Bristol
Year 2023 Dataset release year
Modalities Audio | Video Data modalities included in the dataset
Collection name EPIC-KITCHENS Common name for all related datasets, used to group datasets coming from same source
Research domain SED Strong annotation Related domains, e.g., Scenes, Mobile devices, Audio-visual, Open set, Ambient noise, Unlabelled, Multiple sensors, SED, SELD, Tagging, FL, Strong annotation, Weak annotation, Unlabelled, Multi-annotator
Related datasets name
License Creative Commons, CC BY-NC 4.0
Download Download
Companion site Site Link to the companion site for the dataset
Citation [Huh2023] Epic-Sounds: A Large-scale Dataset of Actions That Sound

Audio

Label Value Description

Data

Data type Audio Possible values: Audio | Features
File format
File format type Constant Possible values: Constant | Variable
Lossy compression No is audio compressed in a lossy manner
Bit rate 16 Bit depth of audio, possible values: 8 | 16 | 24 | 32
Sampling rate (kHz) 24 kHz Sampling rate in kHz, possible values: 8 | 16 | 22.05 | 32 | 44.1 | 48
Channels
Setup Mono Possible values: Mono | Stereo | Binaural | Ambisonic | Array | Multi-Channel | Variable
Number of channels 1
Material
Source EPIC-KITCHENS-100 Possible values: Original | Youtube | Freesound | Online | Crowdsourced | [Dataset name]

Content

Content type Freefield Possible values: Freefield | Synthetic | Isolated
Scene Variable Is the scene class constant for single recording, possible values: Constant | Variable
Event / Spatial location Moving Possible values: Constant | Moving | Unknown

Recording

Setup Uncontrolled Possible values: Near-field | Far-field | Mixed | Uncontrolled | Unknown
Setup count 1 Amount of different recording setups (microphone and recording device) used
Spot type Unknown Possible values: Fixed | Moving | Unknown

Files

Count 700 files Total number of files
Total duration (minutes) 6002 min Total duration of the dataset in minutes
File length Variable Characterization of the file lengths, possible values: Constant | Quasi-constant | Variable
File length (seconds) 10-3708 sec Approximate length of files

Meta

Label Value Description
Types Event Timestamp List of meta data types provided for the data, possible values: Event, Tag, Scene, Caption, Geolocation, Spatial location, Annotator, Timestamp, Presence, Proximity, etc.

Event

Classes 44 Number of event classes
Classes False Possible values: True | False | Almost
Classes
  • background
  • beep
  • ceramic / glass collision
  • ceramic / marble collision
  • ceramic / wood collision
  • ceramic-only collision
  • click
  • cloth-only collision
  • cut / chop
  • drink / eat
  • footstep
  • glass / marble collision
  • glass-only collision
  • hoover / fan
  • human
  • kettle / mixer / appliance
  • kneading
  • metal / ceramic collision
  • metal / cloth collision
  • metal / glass collision
  • metal / marble collision
  • metal / paper collision
  • metal / plastic collision
  • metal / wood collision
  • metal-only collision
  • open / close
  • paper-only collision
  • plastic / ceramic collision
  • plastic / glass collision
  • plastic / marble collision
  • plastic / paper collision
  • plastic / wood collision
  • plastic-only collision
  • pour
  • rustle
  • scrub / scrape / scour / wipe
  • sizzling / boiling
  • slide object
  • spray
  • stir / mix / whisk
  • water
  • wood / glass collision
  • wood-only collision
  • zip
Annotation
Type Strong Possible values: Strong | Weak | Location | None
Annotations per item 1 How many annotations there are available per item (possible multi-annotator setup)
Labelled amount (%) 66.7 % Percentage of all data, amount of data which is labelled
Validated amount (%) 66.7 % Percentage of all data, amount of data which is validated by human
Strong annotations amount (%) 66.7 % Percentage of all data, amount of data which has strong annotations
Overlapping event instances Yes
Labeling
Hierarchical No
Ontology name Yes
Instance
Count 117553 Count of all event instances in the dataset
Average instances per class 2672 Average per class instance count

Cross-validation setup

Label Value Description
Provided Yes
Sets Train Val Test Set types provided in the split, possible values: Train | Test | Val | Dev | Eval

Baseline

Label Value Description
Download Download Link to baseline system source code
Citation [Huh2023] Paper to cite for the baseline