Filterbank-Inspired Neural Analysis–Synthesis

Diagram of MRDLA multiresolution analysis and synthesis for source separation

We investigate filterbank-inspired neural analysis–synthesis structures, including their application to audio source separation. Multiresolution Deep Layered Analysis (MRDLA) draws on wavelet-inspired multiresolution analysis. A distinct approach uses orthogonal convolutional layers to construct invertible analysis–masking–synthesis networks. The invertibility guarantee applies to the orthogonal-convolution approach, not to MRDLA.

Demo

Related Publications

Ren Uchida, Kohei Yatabe, and Tomohiko Nakamura, “Encoder-masking-decoder networks using orthogonal convolutional layer as invertible linear encoder,” Acoustical Science and Technology, vol. advpub, no. e26.10, May 2026.
Tomohiko Nakamura, Shihori Kozuka, and Hiroshi Saruwatari, “Time-domain audio source separation with neural networks based on multiresolution analysis,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 1687–1701, Apr. 2021.
Shihori Kozuka, Tomohiko Nakamura, and Hiroshi Saruwatari, “Investigation on wavelet basis function of DNN-based time domain audio source separation inspired by multiresolution analysis,” in Proceedings of International Congress and Exposition on Noise Control Engineering, Aug. 2020, pp. 4013–4022.
Tomohiko Nakamura and Hiroshi Saruwatari, “Time-domain audio source separation based on Wave-U-Net combined with discrete wavelet transform,” in Proceedings of IEEE International Conference on Acoustics, Speech, and Signal Processing, May 2020, pp. 386–390.