Potential of Non-Uniformly Partitioned Convolution with Freely Adaptable FFT Sizes

Wefers F, Vorländer M. Potential of Non-Uniformly Partitioned Convolution with Freely Adaptable FFT Sizes. In: AES Convention 133. Audio Engineering Society; 2012. Paper 8788. Available from: https://aes.org/publications/elibrary-page/?id=16530

@inproceedings{Wefers2012_16530,
  author = {Wefers, Frank and Vorländer, Michael},
  title = {{Potential of Non-Uniformly Partitioned Convolution with Freely Adaptable FFT Sizes}},
  booktitle = {AES Convention 133},
  note = {Paper 8788},
  year = {2012},
  month = oct,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=16530}
}

TY  - CPAPER
TI  - Potential of Non-Uniformly Partitioned Convolution with Freely Adaptable FFT Sizes
AU  - Wefers, Frank
AU  - Vorländer, Michael
T2  - AES Convention 133
M1  - Paper 8788
PY  - 2012
DA  - 2012/10/06
UR  - https://aes.org/publications/elibrary-page/?id=16530
PB  - Audio Engineering Society
LA  - en
AB  - The standard class of algorithms used for FIR filtering with long impulse responses and short input-to-output latencies are non-uniformly partitioned fast convolution methods. Here a filter impulse response is split into several smaller sub filters of different sizes. Small sub filters are needed for a low latency, whereas long filter parts allow for more computational efficiency. Finding an optimal filter partition that minimizes the computational cost is not trivial, however optimization algorithms are known. Mostly the Fast Fourier Transform (FFT) is used for implementing the fast convolution of the sub filters. Usually the FFT transform sizes are chosen to be powers of two, which has a direct effect on the partitioning of filters. Recent studies reveal that the use of FFT transform sizes that are not powers two has a strong potential to lower the computational costs of the convolution even more. This paper presents a new real-time low-latency convolution algorithm, which performs non-uniformly partitioned convolution with freely adaptable FFT sizes. Alongside, an optimization technique is presented that allows adjusting the FFT sizes in order to minimize the computational complexity for this new framework of non-uniform filter partitions. Finally the performance of the algorithm is compared to conventional methods.
ER  -