Haljan Lugo Girao, M.Sc.

Reserach Fields

• Deep Learning
• Acoustic Echo Cancellation

Thesis topics

Comparing Flow Matching and Diffusion Techniques for Speech Enhancement Systems
Flow matching provides a promising generalization of diffusion-based generative approaches. Thus, recent state-of-the-art image generation models motivate the adoption of flow matching for speech enhancement tasks like acoustic echo cancellation or noise reduction. In this work, you will explore flow matching and compare it to already established diffusion-based score matching model formulations.

Finding Content-Adaptive Speech Representations via 2D Gaussians
Inspired by recent advances in image compression that leverage Gaussian splatting to achieve extreme compression ratios without sacrificing visual quality, this work investigates the application of similar techniques to speech. You will explore if these techniques translate to the compression of spectrogram representations and assess whether the resulting reconstructions preserve audio fidelity at competitive compression rates.

Exploring Attention Mechanisms for Acoustic Echo Cancellation
Self- and cross-attention are the backbone of modern LLMs and other sequence-to-sequence models. Especially cross-attention may be highly advantageous to model correspondences of the loudspeaker and microphone signal in acoustic echo control (AEC) tasks. In this work, you will investigate the methods to incorporate these attention mechanisms to state-of-the-art AEC systems.

Measuring Intelligibility in Speech Processing Tasks - On LPS and Beyond
Currently the Levensthein phoneme similarity (LPS) provides the best assessment on intelligibility utilizing the Levenshtein distance between predicted phonemes on both processed and clean speech samples. One possible question to explore in this work is whether one can find robust time aligned representation on speech intelligibility in a similar way.