A novel audio representation using space filling curves
This addresses the challenge of audio representation for deep learning applications, offering an incremental improvement over existing methods.
The paper tackles the problem of representing audio signals for CNNs by mapping 1D audio waveforms to 2D images using space filling curves, preserving local structure without compression, and shows that the Z curve yields the best results in keyword spotting tasks, producing comparable performance to mel frequency cepstral coefficients.
Since convolutional neural networks (CNNs) have revolutionized the image processing field, they have been widely applied in the audio context. A common approach is to convert the one-dimensional audio signal time series to two-dimensional images using a time-frequency decomposition method. Also it is common to discard the phase information. In this paper, we propose to map one-dimensional audio waveforms to two-dimensional images using space filling curves (SFCs). These mappings do not compress the input signal, while preserving its local structure. Moreover, the mappings benefit from progress made in deep learning and the large collection of existing computer vision networks. We test eight SFCs on two keyword spotting problems. We show that the Z curve yields the best results due to its shift equivariance under convolution operations. Additionally, the Z curve produces comparable results to the widely used mel frequency cepstral coefficients across multiple CNNs.