Master's Thesis in Mathematics / Computer Science or CE/EE

05.09.2026, Abschlussarbeiten, Bachelor- und Masterarbeiten

Open thesis project at TUM CIT, supervisor Dr. rer. nat. B. Hoock
If interested, please use the portal at https://thesis.aet.cit.tum.de/topics or directly send your application to benedikt.hoock(at)tum.de.

Topic description:
Using deep learning for audio compression is an active field of research, advancing beyond perceptual coding like MP3. However, the quantitative measure of optimal compression is a non-trivial issue in this field: physical distortion of the original waveforms is not the same as perceived distortion, and it is the latter that actually defines an optimal neural coder. The goal of this Master's thesis is to find, analyze, and compare different definitions of loss functions that define the perceived loss in a mathematically rigorous way. Specifically, these loss functions should be differentiable to work well with backpropagation in neural network training.

Requirements:
- The topic is best suited for students of mathematics, electrical engineering, "Elektrotechnik und Informationstechnik", communications and electronics engineering, and information engineering.

- solid understanding of signal processing, especially Fourier transform

- solid understanding of deep learning / machine learning

- programming skills in a deep learning framework (PyTorch or TensorFlow preferred)

- ideally, basic knowledge in audio processing

- also welcomed is playing an instrument or engaging in music as a hobby


Goals:

- Review of state-of-the-art neural sound compression

- Mathematical formulations of perceived distortion as a differentiable loss function

- Numerical testing and comparison of the loss functions


References:

- Zölzer, U., 2022. Digital audio signal processing. John Wiley & Sons.

- Cui, W., Yu, D., Jiao, X., Meng, Z., Zhang, G., Wang, Q., Guo, S.Y. and King, I., 2025, July. Recent advances in speech language models: A survey. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 13943-13970).

- Wu, H., Chen, X., Lin, Y.C., Chang, K., Du, J., Lu, K.H., Liu, A.H., Chung, H.L., Wu, Y.K., Yang, D. and Liu, S., 2024, December. Codec-superb@ slt 2024: A lightweight benchmark for neural audio codec models. In 2024 IEEE Spoken Language Technology Workshop (SLT) (pp. 570-577). IEEE.

- Shin, S., Byun, J., Park, Y., Sung, J. and Beack, S., 2022, May. Deep neural network (DNN) audio coder using a perceptually improved training method. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 871-875). IEEE.

(meant as a start for snowballing)

Kontakt: benedikt.hoock@tum.de