Doctoral theses of the School of Electrical Engineering at Aaltodoc (external link)
Doctoral theses of the School of Electrical Engineering are available in the open access repository maintained by Aalto, Aaltodoc.
Title of the thesis: Spatial Post-Filtering for Speech Enhancement and Source Separation
Thesis defender: Stefan Wirler
Opponent: Prof. Nils Peters, Trinity College Dublin, Ireland
Custos: Prof. Ville Pulkki, Aalto University School of Electrical Engineering
Microphone arrays enable separation by beamforming of sound scenes and are central to applications such as speech enhancement in meetings, teleconferencing, and assistive listening. In practice, however, strong interference, and background noise, especially in systems with limited resources, still remains challenging. To further improve the performance in terms of noise suppression, post-filters can be applied to microphone signals or beamformer outputs. Moreover, spatial post-filters that exploit directional signals can further increase spatial selectivity on top of beamforming.
This thesis introduces spatial post-filters that improve interference suppression for low-order and distributed microphone setups. It shows that the Cross-Pattern Coherence (CroPaC) post-filter can be estimated from an omnidirectional signal and a first-order directional signal, enabling spatial filtering with fewer microphones. A space-domain formulation (SD-CroPaC) is used for linear arrays, including a dual-line configuration and multiplicative post-filter combinations to improve selectivity and reduce common linear-array ambiguities. The thesis further proposes non-linear combinations of low-order CroPaC post-filters for stronger separation with only first-order patterns. Finally, it introduces a distributed-microphone post-filter that aggregates phase-corrected cross-spectral similarities across microphone pairs and fuses the resulting pairwise post-filters into a single overall post-filter.
Overall, the results demonstrate that the proposed post-filters can improve interference suppression of speech sources and remain robust in reverberant conditions across a range of array configurations, including compact low-order directional setups, linear arrays, and widely distributed microphone placements.
Thesis available for public display 7 days prior to the defence at Aalto University's public display page.
Doctoral theses of the School of Electrical Engineering are available in the open access repository maintained by Aalto, Aaltodoc.
The Aalto Acoustics Lab is a multidisciplinary research center focusing on audio processing and spatial sound technologies. The laboratory gathers professors and research teams from three different units: Department of Information and Communications Engineering, Department of Computer Science, and Department of Art and Media.