Masked Conditional Neural Networks for Environmental Sound Classification

Abstract

The ConditionaL Neural Network (CLNN) exploits the nature of the temporal sequencing of the sound signal represented in a spectrogram, and its variant the Masked ConditionaL Neural Network (MCLNN) 1 induces the network to learn in frequency bands by embedding a filterbank-like sparseness over the network's links using a binary mask. Additionally, the masking automates the exploration of different feature combinations concurrently analogous to handcrafting the optimum combination of features for a recognition task. We have evaluated the MCLNN performance using the Urbansound8k dataset of environmental sounds. Additionally, we present a collection of manually recorded sounds for rail and road traffic, YorNoise, to investigate the confusion rates among machine generated sounds possessing low-frequency components. MCLNN has achieved competitive results without augmentation and using 12% of the trainable parameters utilized by an equivalent model based on state-of-theart Convolutional Neural Networks on the Urbansound8k. We extended the Ur-bansound8k dataset with YorNoise, where experiments have shown that common tonal properties affect the classification performance.

Recognition of Acoustic Events Using Masked Conditional Neural Networks

“…We used the model specified in Table I and the signal representation (60 mel-spec with delta) discussed earlier, which is the same transformation used by Piczak-CNN [7]. The dataset is pre-distributed into 10-folds, which we used to report the mean accuracy in Table II. The shallow MCLNN in combination with a long segment (k=50) achieved an accuracy of 74.22% compared to a deep MCLNN with a shorter segment (k=5) in [24]. The accuracy of the MCLNN surpasses other reported neural networks based attempts using state-of-the-art CNN architectures proposed by Salamon et al in [31] and Piczak in [7].…”

“…We have performed the MCLNN evaluation using the Urbansound8k [28], YorNoise [24], ESC-10 [29] and ESC-50 [29] environmental sound datasets. We will discuss the composition of each dataset with the common preprocessing applied, and we will defer the discussion to each dataset's relevant section.…”

“…We will discuss the composition of each dataset with the common preprocessing applied, and we will defer the discussion to each dataset's relevant section. In this work, we explore the performance of a shallow architecture of the MCLNN in combination with a long segment compared to the deep MCLNN architectures considered for the mentioned datasets in [23][24][25].…”

“…Table III lists the mean accuracies achieved over a 10-fold cross-validation for both the Urbansound8k and YorNoise combined. A shallow MCLNN achieved an accuracy of 75.92% compared to the deep architecture in [24] that reached 75.13%. Despite the comparable accuracy, the shallow MCLNN used 1 million parameters compared to the 3 million parameters of the deep variant and achieved higher accuracy using a longer segment.…”

“…Additionally, Piczak used augmentation, which involves introducing deformations to sound signal, e.g. time delay, pitch [24] 75.13 shifting. Piczak applied 10 augmentation variants to each sound file, which increases the dataset and consequently the accuracy as studied by Salamon in [31].…”