A Sequential Update Algorithm with Parallel Processing for Deep Learning-Based Beamforming in Real Environments
Sungwook Yoon, Oh‐Wook Kwon · 2021 International Conference on Information and Communication Technology Convergence (ICTC) · 2021
In this paper, we propose a deep learning-based sequential update beamforming algorithm for continuous speech enhancement in real environments. Previous beamforming algorithms have evaluated the performance using pre-segmented audio signals generated by mixing speech and noise in a completely overlapped manner. However, in real environments, speech utterances are sparsely uttered on the time axis and then input signals of beamformers become a continuous noisy speech stream with no speech most of the time. When a noise signal without speech is input to the previous beamforming algorithms, the performance is degraded. Considering these facts, we propose a robust block-wise sequential update algorithm in the real environment. In addition, by using parallel processing, the convergence speed of the update algorithm was increased, resulting in additional performance improvement. The evaluation set is composed of a mixture of speech utterances extracted from the CHiME3 evaluation set and the continuously-played background music extracted from MUSDB. As a speech recognizer, the Kaldi-based toolkit is used. We confirm that the proposed sequential update beamforming algorithm shows better performance than the baseline beamforming algorithm.