Direction-aware target speaker extraction with a dual-channel system based on conditional variational autoencoders under underdetermined conditions

In this paper, we deal with a dual-channel target speaker extraction (TSE) problem under underdetermined conditions. For the dual-channel system, the generalized sidelobe canceller (GSC) is a commonly used structure for estimating a blocking matrix (BM) to generate interference, and geometric source separation (GSS) can be used as an implementation of BM estimation utilizing directional information. However, the performance of the conventional GSS methods is limited under underdetermined conditions because of the lack of a powerful source model. In this paper, we propose a dual-channel TSE method that combines the ability of target selection based on geometric constraints, more powerful source modeling, and nonlinear postprocessing. The target directional information is used as a geometric constraint, and two conditional variational autoencoders (CVAEs) are used to model a single speaker's speech and interference mixture speech. For the postprocessing, an ideal ratio Time-Frequency (T-F) mask estimated from the separated interference mixture speech is used to extract the target speaker's speech. The experimental results demonstrate that the proposed method achieves 6.24 dB and 8.37 dB improvements compared with the baseline method in terms of signal-to-distortions ratio (SDR) and source-to-interferences ratio (SIR) respectively under strong reverberation for 470 ms.