An Unsupervised Deep Learning System for Acoustic Scene Analysis

Abstract

Acoustic scene analysis has attracted a lot of attention recently. Existing methods are mostly supervised, which requires well-predefined acoustic scene categories and accurate labels. In practice, there exists a large amount of unlabeled audio data, but labeling large-scale data is not only costly but also time-consuming. Unsupervised acoustic scene analysis on the other hand does not require manual labeling but is known to have significantly lower performance and therefore has not been well explored. In this paper, a new unsupervised method based on deep auto-encoder networks and spectral clustering is proposed. It first extracts a bottleneck feature from the original acoustic feature of audio clips by an auto-encoder network, and then employs spectral clustering to further reduce the noise and unrelated information in the bottleneck feature. Finally, it conducts hierarchical clustering on the low-dimensional output of the spectral clustering. To fully utilize the spatial information of stereo audio, we further apply the binaural representation and conduct joint clustering on that. To the best of our knowledge, this is the first time that a binaural representation is being used in unsupervised learning. Experimental results show that the proposed method outperforms the state-of-the-art competing methods.

One‐dimensional convolutional neural networks for high‐resolution range profile recognition via adaptively feature recalibrating and automatically channel pruning

“…Thus each channel yields a feature map where denotes the convolution operator, denotes the bias for th feature map, and is a nonlinear activation function. In this paper, Mish 42 function is applied to the 1D‐CNN as the activation function, which is defined as follows: where refers to the softplus 43 activation. More specially, the value of the conv1d operation at position is given by where refers to Hadamard product and is the window size of filters in the th conv1d layer.…”

“…denotes the convolution operator, b k l ( , ) denotes the bias for kth feature map, and δ (•) is a nonlinear activation function. In this paper, Mish 42 function is applied to the 1D-CNN as the activation function, which is defined as follows:…”

Binaural Modelling and Spatial Auditory Cue Analysis of 3D-Printed Ears

“…While reproducing the auditory model of binaural hearing may be a challenging problem, the past decade has seen a renewed interest in binaural approaches to sound localization, which has been applied in a wide area of research and development, including rescue and surveillance robots, animal acoustics, as well as human robot interactions [ 29 , 30 , 31 , 32 , 33 , 34 , 35 ]. Unique or predetermined sound sources for instance can be embedded with search and rescue robots for ad-hoc localization in hazardous or cluttered environments, as well as for emergency signals in remote or unknown areas [ 36 , 37 ].…”