Source-linked AI summary

General-purpose Tagging of Freesound Audio with AudioSet Labels: Task Description, Dataset, and Baseline

Eduardo Fonseca, Manoj Plakal, Frederic Font, Daniel P. W. Ellis, Xavier Favory, Jordi Pons, Xavier Serra

arXiv:1807.09902v3cs.SDcs.LGeess.ASstat.ML

TL;DR

The paper addresses general-purpose audio tagging across 41 diverse AudioSet categories using heterogeneous user-generated Freesound audio and annotations of varying reliability. It describes the task, constructs the FSDKaggle2018 dataset, and provides a baseline achieving 0.70 mAP@3 on the entire test set.

  • Problem

    The task targets robust recognition of a broad range of sound events in realistic user-generated audio, including 41 diverse categories and training annotations with varying reliability.

  • Method

    The paper defines a multiclass audio-tagging task, prepares the FSDKaggle2018 Freesound dataset, and separates manually verified from non-verified annotations for baseline development.

  • Results

    0.70 mAP@3 was achieved on the entire test set after five training epochs.

  • Takeaways & Limitations

    The task and dataset provide a benchmark for broadly applicable sound event classifiers using heterogeneous Freesound content and annotations with different reliability levels.

  • Takeaways & Limitations

    Automatically generated candidate annotations depend on user tags and AudioSet mapping, so their reliability depends on subsequent validation.

Abstract

from arXiv · show

This paper describes Task 2 of the DCASE 2018 Challenge, titled "General-purpose audio tagging of Freesound content with AudioSet labels". This task was hosted on the Kaggle platform as "Freesound General-Purpose Audio Tagging Challenge". The goal of the task is to build an audio tagging system that can recognize the category of an audio clip from a subset of 41 diverse categories drawn from the AudioSet Ontology. We present the task, the dataset prepared for the competition, and a baseline system.

1. INTRODUCTION

The paper introduces a general-purpose audio-tagging task spanning 41 diverse AudioSet categories and user-generated Freesound audio. It targets both broad sound-event recognition and training with annotations of varying reliability.

  • Earlier DCASE audio-tagging tasks focused on domestic sounds and smart cars, whereas this task targets general-purpose audio.
  • The task classifies audio clips into one of 41 diverse categories drawn from the AudioSet Ontology.
  • Freesound user-generated audio offers a large research resource but introduces annotation challenges not addressed by earlier DCASE tasks.
  • The task addresses recognizing more diverse sound events and leveraging training subsets with annotations of varying reliability.
  • The resulting systems may support automatic multimedia description and acoustic monitoring applications.

2. TASK SETUP

DCASE 2018 Task 2 is a Kaggle multiclass, single-tag audio-tagging competition that predicts one category per test clip. Evaluation uses ranked predictions and separate public and private leaderboard divisions.

  • The task is a multiclass problem requiring one predicted category for each audio clip in the test set.
  • The competition ran on Kaggle from March 30th to July 31st 2018 under the name “Freesound General-Purpose Audio Tagging Challenge”.
  • The unreleased-label test set was divided into 19% for the public leaderboard and 81% for the private leaderboard.
  • mAP@3 rewards correct labels most when ranked first and accepts up to three predicted labels per clip.
  • Participants submitted predicted labels in CSV files, with at most two submissions daily and two final submissions for ranking.
  • A complementary Judges’ Award recognized novel, problem-specific, and efficient approaches beyond leaderboard performance.

3. DATASET

FSDKaggle2018 is a curated Freesound dataset organized around AudioSet categories, combining manually verified and non-verified annotations. Its construction preserves variable clip durations and annotation reliability while producing a 41-category benchmark.

  • Dataset source: FSDKaggle2018 is a reduced subset of FSD, prepared for the task from Freesound audio content.
  • Dataset source: Freesound provides heterogeneous user-uploaded recordings with tags, titles, descriptions, and Creative Commons licensing.
  • Annotation creation: 268,261 Freesound clips were mapped to AudioSet categories through user-tag mappings, generating weak clip-level candidate annotations.
  • Annotation creation: Candidate annotations were manually validated through user ratings, training examples, category descriptions, quality controls, and agreement-based prioritization.
  • Dataset curation: Manually-verified annotations contain only PP ratings, while non-verified annotations mainly comprise unrated candidates plus a small amount of rated data.
  • Dataset curation: Curation filtered categories and clips by AudioSet hierarchy, duration, licensing, annotation counts, and annotation quality before forming the final dataset.
  • Dataset description: The dataset contains 11,073 mono 44.1 kHz PCM files across 41 categories, with clips lasting 300ms to 30s and nearly 18h of training audio.
  • Dataset limitations: Non-verified clips may contain multiple sound sources despite having one ground-truth label, and their reliability depends on user tags and ontology mapping.

4. BASELINE SYSTEM

The baseline applies a shallow CNN to logmel spectrograms and aggregates overlapping window predictions into clip-level outputs. It reaches an mAP@3 of 0.70 on the entire test set after five training epochs.

  • The baseline uses a relatively shallow 3-layer CNN with logmel spectrogram inputs and a 41-way softmax classifier.
  • Audio is processed as 0.25s overlapping windows with 0.125s hops, using 25ms STFT windows every 10ms and 64 mel bins.
  • 0.70 mAP@3 is achieved on the entire test set after training for 5 epochs.
  • The baseline scores 0.70 on the public leaderboard split and 0.69 on the private split.

5. CONCLUSION

The paper describes the DCASE 2018 Task 2 setup, FSDKaggle2018 dataset, and baseline for general-purpose sound event recognition. The dataset combines varying annotation reliability and variable-length clips to support open research in sound event recognition.

  • The paper presents the task setup, FSDKaggle2018 dataset, and baseline for DCASE 2018 Task 2.
  • FSDKaggle2018 includes training subsets with annotations of varying reliability and variable-length audio clips.
  • The competition and baseline are intended to foster open research in sound event recognition.
Loading 1807.09902v3…