Efficient Low-latency Convolution with Uniform Filter Partition and Its Evaluation on Real-time Blind Source Separation
Yui Kuriki, Taishi Nakashima, Kouei Yamaoka, Natsuki Ueno, Yukoh Wakabayashi, Nobutaka Ono, Sato Ryo · 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) · 2022
In this paper, we discuss an efficient way to realize a low-latency convolution for real-time blind source separation (BSS). In some real-time applications, such as a hearing aid, both low complexity and low latency are necessary. To reduce the computation for the low-latency convolution, a partitioned convolution has been studied. It partitions a filter into multiple blocks, convolves each block with a signal via the frequency domain, and applies the overlap-add. In this paper, we focus on uniform partitioning as a suitable way and introduce it into real-time BSS. The complexity is estimated as the number of multiplications and evaluated with actual implementation on a Raspberry Pi4B. The experimental results indicate that this approach can reduce the execution time of convolution calculation, and the optimum is obtained at a non-trivial block length.