Blind Extraction Of Target Speech Source: Three Ways Of Guidance Exploiting Supervised Speaker Embeddings

The manuscript deals with the robust extraction of a speaker of interest (SOI) from a mixture of audio sources. A blind algorithm based on independent vector extraction (IVE) is used, which, by definition, extracts an arbitrary source. To focus the extraction towards the SOI, a prior knowledge identifying the target source is required. To this end, the manuscript exploits speaker-identification based on embedding features computed via a pretrained forward sequential memory network (FSMN). We introduce and experimentally validate three ways how this prior knowledge can be employed in a blind algorithm, namely, speaker-specific initialization, pilot signal, and supervised deflation of the mixture. The experiments show that the proposed techniques complement each other and lead to robust identification/extraction of the SOI in difficult mixtures of three speakers.