Exploiting Discrete Cosine Transform Features in Speech Enhancement Technique FullSubNet+

The highly effective deep learning-based technique FullSubNet+ employs a full-band and sub-band fusion model to fulfill the speech enhancement task. FullSubNet+ exploits the short-time magnitude spectrogram, real-and imaginary parts of the complex-valued spectrogram to learn the deep neural network that mainly comprises multi-scale time-sensitive channel attention (MulCA) modules and stacked temporal convolution network (TCN) blocks. To capture the phase information of input time-domain signals more simply, we propose using the short-time DCT-based spectrogram as an alternative for the real and imaginary spectrograms to be an input source to learn the FullSubNet+ framework. The preliminary experiments conducted with the VoiceBank-DEMAND task indicate that exploiting STDCT spectrograms in FullSubNet+ achieves higher objective speech quality and intelligibility in terms of PESQ and STOI metric scores, respectively, for the test set compared with the original FullSubNet+ arrangement. In addition, the STDCT-wise FullSubNet+ obtains a real-time factor (RTF) of 0.229, lower than 0.260, the RTF for the original FullSubNet+.