HAL will be down for maintenance from Friday, June 10 at 4pm through Monday, June 13 at 9am. More information
Skip to Main content Skip to Navigation
Preprints, Working Papers, ...

Analysis of weak labels for sound event tagging

Nicolas Turpault 1 Romain Serizel 1 Emmanuel Vincent 1
1 MULTISPEECH - Speech Modeling for Facilitating Oral-Based Communication
Inria Nancy - Grand Est, LORIA - NLPKD - Department of Natural Language Processing & Knowledge Discovery
Abstract : Weak labels are a recurring problem in the context of ambient sound analysis. While multiple methods using neural networks have been proposed to address it, limited attention has been given to the analysis of the problem to have a better understanding of it. Many of these methods seem to improve detection or tagging performance, but they have been evaluated in scenarios where other problems such as unreliable labels, overlapping sound events, or class unbalance also occur. Therefore, it is difficult to conclude whether the observed improvement is due to solving the problem of weak labels or not. In this article, we provide for the first time a detailed analysis of the impact of weak labels independently of other problems on a sound event tagging system. We show that, in order to limit the negative impact of weak labels on the performance, the training clips must be at least as long as the test clips and longer training clip durations have a minor impact. We also show that good temporal aggregation can help to reduce this impact at test time and provide insight on the annotation granularity needed depending on the targeted scenario.
Document type :
Preprints, Working Papers, ...
Complete list of metadata

https://hal.inria.fr/hal-03203692
Contributor : Nicolas Turpault Connect in order to contact the contributor
Submitted on : Wednesday, April 21, 2021 - 12:44:02 AM
Last modification on : Friday, February 4, 2022 - 3:34:57 AM
Long-term archiving on: : Thursday, July 22, 2021 - 6:09:04 PM

File

main_journal.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-03203692, version 1

Citation

Nicolas Turpault, Romain Serizel, Emmanuel Vincent. Analysis of weak labels for sound event tagging. 2021. ⟨hal-03203692⟩

Share

Metrics

Record views

137

Files downloads

182