Sound source characterization using estimated signal components

Abstract

In signal analysis, the time-frequency representation of a signal is of most importance as it provides an efficient framework for analyzing time-varying, multi-component signals. Performing the Short-time Fourier Transform (STFT) outputs the power spectrogram which maps a one-dimensional signal in time to a two-dimensional function in time as well as frequency with a trade-off in temporal and spectral resolutions. This paper presents a method in overcoming the energy smearing problem brought about by the uncertainty principle known as the Gabor Limit. Done through the utilization of the phase information found in acoustic signals, a trace of the original spectrogram without the resolution trade-off was successfully obtained using this method. The analysis and characterization of acoustic signatures was then carried out using an unsupervised machine learning algorithm. Achieving a silhouette score of 0.8419, the clustering algorithm was successful in characterizing signal trajectories as tonal, impulse-like as well as noise-like components.