Multichannel Subband-Fullband Gated Convolutional Recurrent Neural Network for Direction-Based Speech Enhancement with Head-Mounted Microphone Arrays
Benjamin Stahl, Alois Sontacchi · 2023
We present a method for direction-based multichannel speech enhancement with head-mounted microphone arrays. To target the task of direction-based enhancement of a single speaker, a direction-dependent feature extraction composed of matched filters and a maximum-directivity beamformer is employed. The extracted features are then first processed by a subband neural network with an inplace gated convolutional recurrent architecture. In a second stage, a fullband gated convolutional recurrent neural network processes the input features along with the output of the subband network. The proposed method outperforms subband-only and fullband-only approaches in PESQ and SI-SDR on the development datasets of the 2022/23 SPEAR challenge and was rated best among the submissions to the challenge in a paired-comparison listening experiment.