Theses and Student Jobs

General

  • Please contact Prof. Dr. Nils Peters for possibilities regarding theses, research internships, and student jobs.

Specific Offers

Development and Evaluation of a Rotating Microphone

  • Status: Filled
  • Type: Research internship, student job, MSc thesis
  • Description: The goals is to develop a microphone apparatus that rotates with a certain velocity on a fixed circular trajectory. The accurate estimation of the instantaneous microphone position is crucial. A functional prototype will be part of future research projects.
  • Skills: Experience and interest in mechanics, electronics, and python software development.
  • Contact: Prof. Dr. Nils Peters

Learnable acoustical frontend for noise suppression

  • Status: Filled
  • Type: Research internship, MSc thesis
  • Description: In the area of neural processing for audio signals, very often a fixed acoustical frontend extracts features (e.g., bark-scaled spectrograms, MFCCs, or wavelet coefficients). Replacing such static feature extraction process with trainable equivalents let you gain an insight into their working and can produce interesting results. Your task will be finding fixed parameters which can be updated during training. The goal of your evaluation will be noise suppression. RNNoise flavored architecture gives you a starting point where you can make experiments with the acoustical frontend.
  • Skills: some experience in machine learning, background in noise suppression, some knowledge of wavelets is beneficial
  • Contact: Prof. Dr. Nils Peters

A study on affordable neural networks

  • Status: Filled
  • Type: Research internship, MSc thesis
  • Description: In the last decade neural approaches delivered top performing models in many audio fields. Two prominent examples are keyword detection in real environments and scene classification. A main objective for model design is complexity. A more regularized model gives better generalization ability and avoids very complex answers which may turn out to be wrong during testing. The focus of this work are regularization techniques reducing computational complexity as well, making those models more feasible for embedded devices. Already existing techniques are the LassoNet, top-k ReLU, fixed sparsity, matrix factorization, weight quantization and mixture of experts. You will be given two baseline models for keyword detection and scene classification. Together with your supervisor you will develop an overview and make a comparative analysis of mentioned techniques. In a second step the study should compare the trade-off between complexity and performance for selected methods.
  • Skills: experience in machine learning, interest to work with Tensorflow lite/micro or Pytorch Mobile
  • Contact: Prof. Dr. Nils Peters