多チャネル低ランク・スパース分解に基づく柔軟索状レスキューロボットのためのリアルタイム音声強調

抄録

This paper presents a real-time human-voice enhancement method for a hose-shaped rescue robot based on multi-channel low-rank sparse decomposition. Although microphone arrays equipped on hose-shaped robots are crucial for finding victims under collapsed buildings, human voices captured by the microphone array are contaminated by environment-dependent and non-stationary ego-noise. Our method decomposes multi-channel amplitude spectrograms into sparse and low-rank components (human voice and noise) without any prior training. This decomposition is conducted with a state-space model representing the dynamics of these components in a mini-batch manner. Experimental results show that the performance difference between our method and its offline version is less than 3dB in signal-to-distortion ratio.