Noise Reduction using Neural Network Trained with Amplitude Spectra

Noise reduction has been one of the most important issues in acoustic signal processing. There are a wide variety of techniques to enhance the target signal and reduce interference. In recent days, neural network-based approaches achieve a great noise reduction performance. The authors have also proposed a neural network-based beamformer, that is, an autoencoder in the temporal domain. Channel-dependent waveforms obtained by an 8-ch linear microphone array were prepared as training data. It succeeds in forming a sharp main-lobe and suppressing side-lobes when the interference comes from the direction away from the target direction. Its performance drops when the direction of the interference is near that of the target signal. In this paper, the neural network-based beamformer is trained with amplitude spectra of the harmonic signals such as musical instruments. In other words, phase information is not used for noise reduction. The performance of the proposed method is evaluated based on signal-to-noise ratio and spectral distortion. The proposed method has an advantage in noise reduction for interferences, of which arrival directions are close to the target direction.