Active Learning for Auditory Hierarchy - Machine Learning and Knowledge Extraction Access content directly
Conference Papers Year : 2020

Active Learning for Auditory Hierarchy

Abstract

Much audio content today is rendered as a static stereo mix: fundamentally a fixed single entity. Object-based audio envisages the delivery of sound content using a collection of individual sound ‘objects’ controlled by accompanying metadata. This offers potential for audio to be delivered in a dynamic manner providing enhanced audio for consumers. One example of such treatment is the concept of applying varying levels of data compression to sound objects thereby reducing the volume of data to be transmitted in limited bandwidth situations. This application motivates the ability to accurately classify objects in terms of their ‘hierarchy’. That is, whether or not an object is a foreground sound, which should be reproduced at full quality if possible, or a background sound, which can be heavily compressed without causing a deterioration in the listening experience. Lack of suitably labelled data is an acknowledged problem in the domain. Active Learning is a method that can greatly reduce the manual effort required to label a large corpus by identifying the most effective instances to train a model to high accuracy levels. This paper compares a number of Active Learning methods to investigate which is most effective in the context of a hierarchical labelling task on an audio dataset. Results show that the number of manual labels required can be reduced to 1.7% of the total dataset while still retaining high prediction accuracy.
Fichier principal
Vignette du fichier
497121_1_En_20_Chapter.pdf (345.33 Ko) Télécharger le fichier
Origin : Files produced by the author(s)

Dates and versions

hal-03414715 , version 1 (04-11-2021)

Licence

Attribution

Identifiers

Cite

William Coleman, Charlie Cullen, Ming Yan, Sarah Jane Delany. Active Learning for Auditory Hierarchy. 4th International Cross-Domain Conference for Machine Learning and Knowledge Extraction (CD-MAKE), Aug 2020, Dublin, Ireland. pp.365-384, ⟨10.1007/978-3-030-57321-8_20⟩. ⟨hal-03414715⟩
24 View
26 Download

Altmetric

Share

Gmail Facebook X LinkedIn More