Acoustic source separation and localization using UMA-16 microphone array

Abstract

This work addresses blind source separation and direction of arrival estimation in multispeaker acoustic environment. Building on a frequency-domain Independent Vector Analysis framework, a real-time C++ engine was implemented supporting offline batch and true online streaming modes, with hardware interfacing for the MiniDSP UMA-16 array, STFT computation, voice activity detection, and post-filtering. Direction of arrival estimates are derived geometrically from the converged demixing matrix via pseudoinversion. These components are integrated into an interactive PyQt interface overlaying estimated source directions onto a live video stream. Faster-converging variants AuxIVA and OverIVA were additionally evaluated against the NG-IVA baseline.