Time-Domain Target Speaker Extraction with Parallel Intra and Inter-framework

Abstract

Speaker extraction addresses isolating the specific speaker’s voice from a bend of other speakers using supplementary information. This paper proposes a time-domain speaker extraction using a parallel intra- and inter-framework (TSEPII). An efficient intra- and inter-architecture converts mixed utterance into multi-scale embedding coefficients. Additionally, we incorporate parallel architectures to achieve more stability than previous single architectures. This architecture includes the main components such as the auxiliary encoder (the talker encoding block), the extraction encoder (utterance encoding block), the talker extraction block, and the extraction decoder (the utterance decoding block). In particular, the time domain-based raw voice processing system keeps important information. The utterance encoding block transforms the mixed voice into multiple-scale embedding values, while the talker encoding block learns the target talker by the talker embedding feature. The talker extraction block plays an important role and uses multiple-scale embedding values and the talker embedding feature as the input features. It estimates the time-domain mask for the system. Finally, the utterance decoding block recreates the utterance of the target talker. Experiments show that the TSEPII achieves state-of-the-art performance and competes with current methods.