Constant Separating Vector-based Blind Source Extraction and Dereverberation for a Moving Speaker

This paper proposes a multi-channel speech extraction method for moving sound sources in a long reverberant environment. Constant Separating Vector (CSV) mixing model has been devised for batch processing speech extraction to extract a moving target speech stably. Also, based on this mixture model, an update algorithm using auxiliary function technology has been proposed as a fast and stable source extraction. However, source extraction performance will be limited when the reverberation time is long. In recent years, joint optimization technique has been researched to achieve effective dereverberation and source extraction simultaneously under highly reverberant environments. However, the extension to the CSV mixing model is yet to be discovered. To realize moving source extraction under a highly reverberant environment, we derive the update algorithm when the dereverberation mechanism is installed in the conventional method. In our proposed method, we estimate a dereverberation system focusing only on the extracted target sound, which achieves effective source extraction with a small additional computational cost. Our experiment shows that the proposed algorithm achieves sufficient blind dereverberation and source extraction.