Nowadays, time-domain features have been widely used in speech enhancement (SE) networks like frequency-domain features to achieve excellent performance in eliminating noise from input utterances. This study primarily investigates how to extract information from time-domain utterances to create more effective features in speech enhancement. We present employing sub-signals dwelled in multiple acoustic frequency bands in time domain and integrating them into a unified feature set. We propose using the discrete wavelet transform (DWT) to decompose each input frame signal to obtain sub-band signals, and a projection fusion process is performed on these signals to create the ultimate features. The corresponding fusion strategy is the bi-projection fusion (BPF). In short, BPF exploits the sigmoid function to create ratio masks for two feature sources. The concatenation of fused DWT features and time features serves as the encoder output of a celebrated SE framework, fully-convolutional time-domain audio separation network (Conv-TasNet), to estimate the mask and then produce the enhanced time-domain utterances. The evaluation experiments are conducted on the VoiceBank-DEMAND and VoiceBank-QUT tasks. The experimental results reveal that the proposed method achieves higher speech quality and intelligibility than the original Conv-TasNet that uses time features only, indicating that the fusion of DWT features created from the input utterances can benefit time features to learn a superior Conv-TasNet in speech enhancement.

The masking-based speech enhancement method pursues a multiplicative mask that applies to the spectrogram of input noise-corrupted utterance, and a deep neural network (DNN) is often used to learn the mask. In particular, the features commonly used for automatic speech recognition can serve as the input of the DNN to learn the well-behaved mask that significantly reduce the noise distortion of processed utterances. This study proposes to preprocess the input speech features for the ideal ratio mask (IRM)-based DNN by lowpass filtering in order to alleviate the noise components. In particular, we employ the discrete wavelet transform (DWT) to decompose the temporal speech feature sequence and scale down the detail coefficients, which correspond to the high-pass portion of the sequence. Preliminary experiments conducted on a subset of TIMIT corpus reveal that the proposed method can make the resulting IRM achieve higher speech quality and intelligibility for the babble noise-corrupted signals compared with the original IRM, indicating that the lowpass filtered temporal feature sequence can learn a superior IRM network for speech enhancement.