残差駆動型アーキテクチャの提案と音響ストリーム分離への応用

抄録

This paper presents the Residue-Driven Architecture (RDA) as a general computational frame-work for sound stream segregation based on a multi-agent paradigm. Sound stream segregation is an important primary processing for computationally understanding sounds (Computational Auditory Scene Analysis) in the real-world. Since RDA is designed without assuming any specific sound attributes, it can be applied to various kinds of sound stream segregation problems. The RDA consists of three kinds of agents : an event-detector, a tracer-generator, and tracers. The event-detector calculates a residue by subtracting the predicted input from the actual input. When a residue exceeds a threshold value, tracer-generator generates a tracer that extracts a sound stream from the residue and returns a predicted input of the next time frame to the event-detector. The RDA is applied to the design of two subsystems : A monaural subsystem segregates sound streams under background noise using harmonic structure ; a binaural subsystem refines the sound streams segregated by the monaural system using the direction of the sound source. These subsystems can be concisely designed and simply implemented based on the RDA ; therefore, the effectiveness of the RDA is proven. In addition, experimental results show that the capability of the sound stream segregation system is improved by combining these subsystems.