2.4 Function of suppress self-generated sound

In active audition, a sound occurs by creak of a robot itself in some cases as well as the sound of a motor itself occurred by driving the motor. Although the sound that occurs with robot’s movement is small, its sound volume is greater than those of external sound sources according to the inverse-square law since the sound source is near a microphone.

2.4.1 Self-generated sound suppression based on model

Nakadai et al. have attempted to suppress this self-generated sound by setting two microphones in the head of the robot SIG. Having simple templates for motor sounds and machine sounds, when the sound that matches the template occurs during operation of a motor, the subbands that are easily destroyed are broken off with heuristics. The reason why they have used this method is that right and left ears are separately processed in the active noise canceller based on the FIR filter and therefore correct interaural phase differences are not obtained, and furthermore that the FIR filter does not have so much effect on burst noise suppression. Further, in SIG2, microphones are embedded in the model ear canal and a motor used for the robot is silent type, which does not require noise suppression processing. As for QRIO made by Sony, one microphone is embedded in its body and self-generated noise is suppressed by using six microphones aiming at outside. Ince et al. have developed a method to predict self-generated noise caused by the movement of the robot itself by joint angle information and reduce it by the spectral subtraction [12]. Nakadai et al. incorporated a function to reject motor noise from specific directions into HARK [12]. Even et al. have developed a method that estimates directions of the sound radiated from the body surface with three vibration sensors set in the body and adjusts angles of a linear microphone array so that speaker’s direction do not corresponds to the radiated sound’s direction, so as to suppress self-generated sound [12]. For interactions between a robot and human, it is essential to develop a “strategy for better hearing” such as moving to the position where the robot can hear the sound best or turning the body in consideration for influences of the self-generated sound and environment on the sound.

2.4.2 Self-generated sound suppression by semi-blind separation

Figure 2.3: The robot’s own voice enters its own ears accompanied with reverberations and another person’s speech (called barge-in)

In robot audition, it is possible to perform self-generated sound suppression that utilizes the point that self-generated signals are known to the robot itself. Takeda et al., from a semi-blind separation technique based on ICA, have developed a function of self-generated sound suppression that estimates reverberation with self-generated utterances as already-known, suppresses the input sound mixture in self-generated utterances and extracts utterances of a partner in the situation shown in Figure2.3 [12].

Barge-in-capable utterance recognition and a music robot (described later) have been developed as a prototype of the application of this technique. Barge-in-capable utterance is a function that allows humans to speak freely even during the utterance of the robot. When a user barges in and the user utters “that”, “the second one” or “Atom” while a robot provides information enumerating items, the robot can judge which item has been designated from the utterance content and timing with higher accuracy, based on this technology. For symbosis of human beings and robots, it is essential to have mixed-initiative interactions, which allow them to speak freely at any time, not alternating speaking. This self-generated sound suppression method realizes such a function. In the semi-blind separation technique, a self-generated sound enters to ears though it is deleted when separated and therefore it cannot be used for higher-order processing. According to “Brain that speaks and hears words” written by Honjo, in adults, their own voice enters the primary auditory cortex of the temporal lobe, though is not transmitted to the associative auditory cortex of the cerebral cortex, and thus their voice is ignored. The above self-generated sound suppression by the semi-blind separation can be understood as an engineering case of the processing that is terminated at a primary auditory cortex.