Deep-learning-based Single-channel Sound Source Separation in Noisy Environments
Ken’ichi Furuya, Iori Miura · 2024
This study experimentally investigates the effects of stationary noise on separation performance when using single-channel sound source separation based on deep learning in a real environment. Additionally, the differences in the performance of sound source separation models are investigated. The results of the experiment show that the sound source separation model, which was trained on noiseless speech data, exhibits a considerable deterioration in separation performance when stationary noise is introduced. However, using a noise suppression model for preprocessing yields a separation performance that is equivalent to or superior to that of the sound source separation model trained using noisy speech data.