Master Thesis in Extending GEMM for Time-series DSP Algorithms

Prior to feeding data to neural networks, spectrum is typically generated using sliding windows FFT and MFCC on acoustic signal. This approach treats acoustic signal as image and image-based neural networks such as CNN is utilized to perform various tasks such as keyword spotting and denoising.

• Extracting temporal and frequency information using spectrum requires heavy pre-processing due to this approach. In combination with CNN-based neural networks, you will need to develop and optimize the algorithms to utilize the gain provided by the GEMM-based accelerators.
• Here, you will explore different approaches to exploit features presented on acoustic signal. Leveraging time encoding neural networks, time series characteristic of acoustic signal can be better presented with light-weight pre-processing.
• Furthermore, you will investigate various inputs data representation and various DSP-heavy pre- and post-processing to analyze acoustic scenes to allow efficient mapping on the GEMM-based accelerators.
• Finally, hardware design consideration will be the main criteria on designing and optimizing such processing chains that include the design of neural networks to make the hardware implementation feasible.

• Education: Master studies in the field of Electrical Engineering, Computer Science or comparable
• Experience and Knowledge: in Digital Design, (System)Verilog/VHDL, Python; background in Neural Networks
• Personality and Working Practice: you excel at organizing your tasks in a structured manner and working independently
• Enthusiasm: keen interest in future technologies and trends with passion for innovation
• Languages: fluent in English, German is a plus