Improving the wavelet transform based adaptive FullSubNet+ with Huber loss
Pengle Li, Peiran Wu, Jeih-Weih Hung · IET conference proceedings. · 2024
By utilizing the adaptive sub-band encoder, the adaptive FullSubNet+ (A-FSN) is a revision of the well-known speech improvement framework FullSubNet+. This allows the A-FSN to capture the local frequency information in a more flexible manner. In addition, it has been demonstrated that AFSN operates more effectively when its real and imaginary spectrogram inputs are replaced by the short-time discrete wavelet transform (DWT) features. The DWT features preserve the phase information of input time-domain signals in addition to the complex-valued spectrogram, and they also serve as an additional form of time-domain features with fixed kernels to learn the A-FSN framework.