TUT Acoustic scenes 2017, Evaluation & Development datasets, processed image

Unseparated Pulse Energy Spectrogram

Processed audio data.

Sound source separation is a preliminary for acoustic scene classification. It can be argued that rare sound detection can be performed without separation, but in most cases it also depends on it.

I have come up with the theory that the full time-domain, or if assumptions are made on the amplitude-waveform or the phase-profile, even the sequence of events can be discarded for acoustic scene classification.

For short time frame bins, a statistical representation should be enough to correctly identify the scene. Even more so, if deep learning methods are applied.

I have also come up with the theory that energy scalograms are applied pulse-length or waveform/profile-length wise. This can enhance the input representation for machine learning.

Furthermore I have used derivatives of the time signal and applied similar signal processing methods to them. For visualisation I have added them to the original scalogram in different colors. The use of derivatives is very much distorted, if the sound is not separated.