Our purpose is to extract a target signal with binaural microphones in an environment where multiple speech sources exist. A previous system has been established to separate the observed signals and identify which segregation speech is desired. The identification is based on SPDV (Spectral Phase Difference Variance). In this paper, we introduce an auto-correlation-based phase unwrapping to improve the estimation accuracy of SPDV, and an upper limit of the number of analysis frames to reduce the computational cost. Simulations were carried out to confirm the effectiveness of the proposed system. Results show that it works in real-time and gives SDR (Signal-to-Distortion Ratio) slightly superior to conventional source separation techniques which are batch processors.
J-STAGEがリニューアルされました! https://www.jstage.jst.go.jp/browse/-char/ja/
