Target Speech Extraction: Independent Vector Extraction Guided by Supervised Speaker Identification

This manuscript proposes a novel robust procedure for the extraction of speaker interest (SOI) from mixture audio sources. The estimation SOI is performed via independent vector (IVE). Since blind IVE cannot distinguish target source by itself, it guided towards frame-wise identification based on deep learning. Still, an incorrect can be extracted due to guidance failings, especially when processing challenging data. To identify such cases, we propose criterion non-intrusively assessing estimated speaker. It utilizes same model as identification, so no additional training required. When detected, “deflation” step in which subtracted and, subsequently, another attempt extract performed. process repeated until successful achieved. proposed experimentally tested artificial and real-world datasets containing phenomena: movements, reverberation, transient noise, or microphone failures. method compared with state-of-the-art algorithms well current fully supervised learning-based methods.