Automatic speech recognition under robot ego noises

When the robots are moving any part of their body, which inevitably produce noises, such noises are known as the ego noises. These noises caused by various body joint motors or other motors as well as cooling fans for CPU and etc. Moreover, these noises are easily captured by the robots´ microphones, because the noise sources are closer to the microphones than the target speech source. This paper proposes a new framework for de-noising the motor noise. According to noise category, one method of spectral subtraction, joint noise template subtraction, labeled area cepstral mean subtraction and multi-condition training has been selected to suppress and estimate ego noises to improve the performance of automatic speech recognition. Finally, with the ego noises generated by the robot, a series of experimental results prove that our method can significantly reduce the effect of ego-noises and thereby enhance the robustness of automatic speech recognition.