SpeakerBeam: Speaker Aware Neural Network for Target Speaker Extraction in Speech Mixtures

The processing of speech corrupted by interferingoverlapping speakers is one of the challenging problems withregards to todays automatic speech recognition systems. Recently,approaches based on deep learning have made great progresstoward solving this problem. Most of these approaches tacklethe problem as speech separation, i.e., they blindly recover allthe speakers from the mixture. In some scenarios, such as smartpersonal devices, we may however be interested in recovering onetarget speaker froma mixture. In this paper, we introduce Speaker-Beam, a method for extracting a target speaker from the mixturebased on an adaptation utterance spoken by the target speaker.Formulating the problem as speaker extraction avoids certainissues such as label permutation and the need to determine thenumber of speakers in the mixture.With SpeakerBeam, we jointlylearn to extract a representation from the adaptation utterancecharacterizing the target speaker and to use this representationto extract the speaker. We explore several ways to do this, mostlyinspired by speaker adaptation in acoustic models for automaticspeech recognition. We evaluate the performance on the widelyused WSJ0-2mix andWSJ0-3mix datasets, and these datasets modifiedwith more noise or more realistic overlapping patterns. Wefurther analyze the learned behavior by exploring the speaker representationsand assessing the effect of the length of the adaptationdata. The results show the benefit of including speaker informationin the processing and the effectiveness of the proposed method.