Date – Details
Master-Presentation: Learning Latent Representations for Acoustic System Identification on Manifolds
Ysrael Savares Salire
Thursday, April 30, 2026
11:15 AM
IKS 4G | zoom
Distortion-free audio is a fundamental expectation in modern telecommunication. In hands-free situations, the loudspeaker signal may couple into the microphone through electro-acoustic paths, leading to echoes. This is addressed through Acoustic Echo Cancellation (AEC). In a neural network-based approach to AEC, a Variational Autoencoder (VAE) is trained on a dataset of Room Impulse Responses (RIRs). The VAE’s latent space provides a lower-dimensional manifold where nearby points in the latent space correspond with points that are similar in the higher-dimensional observation space. This latent space provides the hidden state of an Extended Kalman Filter (EKF) applied to AEC.
This thesis explores whether regularization strategies that encourage more latent disentanglement, like Disentangled Inferred Prior Variational Autoencoder (DIPVAE) and Total Correlation Variational Autoencoder (TCVAE), lead to improvements in AEC performance over the standard VAE. Latent disentanglement is theoretically advantageous as it aligns with an architectural choice to estimate process noise as diagonal covariance matrices. Performance was evaluated by comparing the measured Echo Return Loss Enhancement (ERLE) over AEC scenarios with different Echo-to-Noise Ratios (ENRs) of speech and noise excitation signals. Primary consideration is given to the MIRaGe dataset which provides loudspeaker coordinates that can be used to calculate disentanglement.
While latent disentanglement scores could not be positively correlated with AEC performance, the tested regularization strategies did otherwise lead to up to 2.5 dB improvement in measured ERLE. Experiments suggested that the AEC benefited from encouraging the VAE to learn a more linear manifold. Other improvements could be seen in increased robustness against adaptive filter stalling and in reducing posterior collapse.
