Low Latency Online Source Separation and Noise Reduction Based on Joint Optimization with Dereverberation
Tetsuya Ueda, Tomohiro Nakatani, Rintaro Ikeshita, Keisuke Kinoshita, Shoko Araki, Shoji Makino · 2021 29th European Signal Processing Conference (EUSIPCO) · 2021
This paper proposes low latency online source separation in noisy environments. An approach based on weighted prediction error dereverberation was recently proposed to solve the degradation caused by using low latency online source separation. Although this approach can also reduce noise by increasing the number of microphones and separating the noise as additional sources, the calculation cost prohibitively increases. To solve this problem, this paper incorporates techniques used in independent vector extraction (IVE) into the above conventional approach. Because IVE can skip most of the calculations for estimating noise by assuming that it is a stationary Gaussian, our proposed method achieves effective and computationally efficient noise reduction using many microphones. Experiments in a noisy car environment show that our proposed online method simultaneously separates sources and reduces noise with low latency (< 12 ms) processing.