Acoustic Scene Classification using Convolutional Neural Networks on Multivariate Audio

Research has shown the efficacy of using convolutional neural networks (CNN) with audio
spectrograms in machine listening tasks such as acoustic scene classification (ASC).
There is, however, a knowledge gap when it comes to standardizing preprocessing practices
for this form of ASC. Researchers using these methods have been moving forward
in relative darkness about how to best represent their audio data for consumption by
a CNN, often relying on transfer learning from adjacent machine listening tasks. This
work explores the relationship of frequency limens and channel depth on ASC accuracy
with CNNs of three different varieties: generic, deep, and wide. Results show
that variability in the representation of spectral audio information plays a crucial role
in classifier performance. Classification accuracy improved when using multi-channel
representations of audio data over a single channel alternative. Classification accuracy
also decreased when the representative spectra contained less frequency information,
albeit to a lesser degree. This pattern was nearly consistent across each of the proposed
CNN architectures. These findings have direct implications for several academic and
industrial machine listening applications. In the academic realm, they work towards
codifying audio data preprocessing practices and network architectural decisions. In
industry, the results open the door for exploring the usage of substandard microphones
in technologies that employ machine listening such as commodity hardware.