In this work, we expand the evaluation of our previously proposed method, a multichannel extension of Language-Queried Audio Source Separation (LASS), to various types of sound sources, including moving sources. This method estimates the source model of Independent Vector Analysis (IVA), combined with language queries, thereby integrating linguistic and spatial information to separate arbitrary sounds. We refer to this method as the LASS model-IVA. While LASS is designed for Universal Source Separation (USS), our previous evaluation was limited to a small number of source types, all of which were located at fixed positions. To explore its applicability in more practical scenarios, we investigate the effectiveness of the LASS model-IVA under more realistic conditions, especially for moving sources. Work supported by JST SICORP Grant No. JPMJSC2306.
Nakamura et al. (2025) studied this question.