Master Thesis Features Exploitation of Acoustic Signals Using Wavelet Networks

Prior to feeding data to neural networks, the spectrum is typically generated using sliding windows FFT and MFCC on acoustic signal. This approach treats the acoustic signal as an image, allowing image-based neural networks, such as CNN, to perform various tasks, including keyword spotting. However, extracting temporal and frequency information from the spectrum requires heavy pre-processing due to this method, and CNN-based neural networks may be ineffective for solving such tasks. During your Master thesis, you will explore various approaches to leverage features present in acoustic signals. By utilizing time-encoding neural networks, the time series characteristics of acoustic signals can be better represented without the need for extensive pre-processing. In our team, you will investigate various inputs data representation methods and network topologies, such as wavelet networks, to analyze acoustic scenes, enabling direct processing of input into neural networks. Additionally, hardware design consideration will be a key factor in designing processing chains, including the design of neural networks, to ensure that the hardware implementation is feasible.