Y. Bando, K. Itoyama, M. Konyo, S. Tadokoro, K. Nakadai, K. Yoshii, and H. G. Okuno
This paper presents an efficient statistical method that enhances human voices included in severely noisy audio signals recorded by microphones of a hose-shaped rescue robot. To help a remote operator of such a robot pick up a weak voice of a human buried under rubble, it is crucial to suppress the loud ego-noise caused by the movements of the robot in real time. A promising approach to this task is to use online robust principal component analysis (ORPCA) for decomposing the spectrogram of an observed noisy signal into the sum of low-rank and sparse spectrograms that are expected to correspond to periodic ego-noise and human voices. Using a microphone array distributed on the long body of a hose-shaped robot, ego-noise suppression can be further improved by combining the results of ORPCA applied to the observed signal captured by each microphone. Experiments using a 3-m hose-shaped rescue robot with eight microphones showed that the performance of ego-noise suppression was improved by 7.4 dB in SDR and 17.2 dB in SIR compared with that of conventional methods using only one microphone.
We estimated the robot postures with following three methods:
HRLE is one of the conventional methods that solve the same noise suppression problem as ORPCA.
[1] J. Fukuda+, “Remote vertical exploration by Active Scope Came rinnto collapsed buildings,” in IEEE/RSJ IROS, 2014, pp. 1882–1888.
[2] J. Feng+, “Online robust PCA via stochastic optimization,” in
Annual Conference on Neural Information Processing Systems, 2013,
pp. 404–412.
[3] H. Nakajima+, “An easily-configurable robot audition system using histogram-based recursive level estimation,” in IEEE/RSJ IROS, 2010, pp. 958–963.