Hypernetworks for Sound event Detection: a Proof-of-Concept

Abstract

Polyphonic sound event detection (SED) involves the pre-diction of sound events present in an audio recording along with their onset and offset times. Recently, Deep Neural Net-works, specifically convolutional recurrent neural networks (CRNN) have achieved impressive results for this task. The convolution part of the architecture is used to extract trans-lational invariant features from the input and the recurrent part learns the underlying temporal relationship between au-dio frames. Recent studies showed that the weight sharing paradigm of recurrent networks might be a hindering factor in certain kinds of time series data, specifically where there is a temporal conditional shift, i.e. the conditional distribution of a label changes across the temporal scale. This warrants a relevant question - is there a similar phenomenon in poly-phonic sound events due to dynamic polyphony level across the temporal axis? In this work, we explore this question and inquire if relaxed weight sharing improves performance of a CRNN for polyphonic SED. We propose to use hyper-networks to relax weight sharing in the recurrent part and show that the CRNN's performance is improved by ≈ 3% across two datasets, thus paving the way for further explo-ration of the existence of temporal conditional shift for poly-phonic SED.