To develop new algorithms for noise reduction/interference cancellation to enhance the beamformer output of the microphone array mounted on the UAV.
The use of acoustic information by mounting microphones on the UAV to pick up signals from the ground level is a challenging task, as the quality of the speech signal has often been reduced due to the rotor noise, the reverberation due to ground reflection, and the distance between the desired source and the UAV.
Some preliminary results showed that traditional methods such as beamforming and post-filtering is able to enhance the speech data. However, due to heavy UAV rotor noise, the signal to noise ratio (SNR) of the speech signal picked up by the microphones mounted on the UAV is very low leading to poor intelligibility of the enhanced signal. Further improvement (25-50%) has been achieved by introducing machine learning technique, where a deep neural network(DNN) is trained using clean speech signal corrupted with UAV noise. Going forwards, the performance and adaptability of the DNN is to be further enhanced by training the DNN with larger datasets including signals recorded on the flight.
Working principle of the proposed speech enhancement algorithm
The performance of the DNN in enhancing speech signal corrupted by the UAV noise. The ideal signal (green) is generated using the ideal mask, whereas the estimated signal (red) is obtained when the mask estimated by the trained DNN is used to recover the signal from the noise signal (black.)