MFFR-net: Multi-scale feature fusion and attentive recalibration network for deep neural speech enhancement

Deep neural networks (DNNs) have been successfully applied in advancing speech enhancement (SE), particularly in overcoming the challenges posed by nonstationary noisy backgrounds. In this context, multi-scale feature fusion and recalibration (MFFR) can improve speech enhancement performance by combining multi-scale and recalibrated features. This paper proposes a speech enhancement system that capitalizes on a large-scale pre-trained model, seamlessly fused with features attentively recalibrated using varying kernel sizes in convolutional layers. This process enables the SE system to capture features across diverse scales, enhancing its overall performance. The proposed SE system uses a transferable features extractor architecture and integrates with multi-scaled attentively recalibrated features. Utilizing 2D-convolutional layers, the convolutional encoder-decoder extracts both local and contextual features from speech signals. To capture long-term temporal dependencies, a bidirectional simple recurrent unit (BSRU) serves as a bottleneck layer positioned between the encoder and decoder. The experiments are conducted on three publicly available datasets including Texas Instruments/Massachusetts Institute of Technology (TIMIT), LibriSpeech, and Voice Cloning Toolkit+Diverse Environments Multi-channel Acoustic Noise Database (VCTK+DEMAND). The experimental results show that the proposed SE system performs better than several recent approaches on the Short-Time Objective Intelligibility (STOI) and Perceptual Evaluation of Speech Quality (PESQ) evaluation metrics. On the TIMIT dataset, the proposed system showcases a considerable improvement in STOI (17.3%) and PESQ (0.74) over the noisy mixture. The evaluation on the LibriSpeech dataset yields results with a 17.6% and 0.87 improvement in STOI and PESQ.