Compact Network for Speakerbeam Target Speaker Extraction

Speech separation that separates a mixture of speech signals intoeach of its sources has been an active research topic for a long timeand has seen recent progress with the advent of deep learning. Arelated problem is target speaker extraction, i.e. extraction of onlyspeech of a target speaker out of a mixture, given characteristics ofhis/her voice. We have recently proposed SpeakerBeam, which isa neural network-based target speaker extraction method. Speaker-Beam uses a speech extraction network that is adapted to the targetspeaker using auxiliary features derived from an adaptation utteranceof that speaker. Initially, we implemented SpeakerBeam with afactorized adaptation layer, which consists of several parallel lineartransformations weighted by weights derived from the auxiliary features.The factorized layer is effective for target speech extraction,but it requires a large number of parameters. In this paper, we proposeto simply scale the activations of a hidden layer of the speechextraction network with weights derived from the auxiliary features.This simpler approach greatly reduces the number of model parametersby up to 60%, making it much more practical, while maintaininga similar level of performance. We tested our approach on simulatedand real noisy and reverberant mixtures, showing the potential ofSpeakerBeam for real-life applications. Moreover, we showed thatspeech extraction performance of SpeakerBeam compares favorablywith that of a state-of-the-art speech separation method with a similarnetwork configuration.