Supervised Representation Learning for Audio Scene Classification

Abstract

Abstract-This paper investigates the use of supervised feature learning approaches for extracting relevant and discriminative features from acoustic scene recordings. Owing to the recent release of open datasets for acoustic scene classification (ASC) problems, representation learning techniques can now be envisioned for solving the problem of feature extraction. This paper makes a step towards this goal by first studying models based on convolutional neural networks (ConvNet). Because the scale of the datasets available is still small compared to those available in computer vision, we also introduce a technical contribution denoted as supervised non-negative matrix factorization (SNMF). Our goal through this SNMF is to induce the matrix decomposition to carry out discriminative information in addition to the usual generative ones. We achieve this objective by augmenting the NMF optimization problem with a novel loss function related to class labels of acoustic scenes. Our experiments show that despite the small-scale setting, supervised feature learning is favorably competitive compared to the current state-of-the-art features. We also point out that for smaller scale dataset, supervised NMF is indeed slightly less prone to overfitting than convolutional neural networks. While the performances of these learned features are interesting per se, a deeper analysis of their behavior in the acoustic scene problem context raises open and difficult questions that we believe, need to be addressed for further performance breakthroughs.

Spectrotemporal Analysis Using Local Binary Pattern Variants for Acoustic Scene Classification

“…The deep learning performance can vary significantly with different feature representation and architecture and a large amount of data is required to train the feature learners [25]. In [9], [26] and [27], it has been demonstrated that the acoustic scenes can be learned from a TFR using matrix factorization techniques. Matrix factorization methods using Principal Component Analysis (PCA) and Nonnegative Matrix Factorization (NMF) have been explored with different variants and tuning strategies to further improve the classification accuracy.…”

Leveraging deep neural networks with nonnegative representations for improved environmental sound classification

“…Therefore, the first step of an ASC system is either to compute a time-frequency representation or hand-crafted features in order to work with more compact and interpretable data. Time frequency representations are mostly used as inputs of matrix factorization or CNN systems [12,13,16]. They are usually based on perceptually motivated time-frequency representations with log-scaled frequency bands such as constant-Q transforms (CQT) or Mel spectrograms.…”

“…The use of other divergences have not shown to provide any notable increase in performance for the task [12] while augmenting the computation time. Supervised NMF models have been applied to ASC with the goal of taking into account the knowledge about the class labels in order to learn better decompositions [12,13,20]. We choose to use the Task-driven NMF (TNMF) formulation [12], a supervised NMF approach adapted from the Taskdriven dictionary learning framework [21].…”

“…The second important trend in ASC is to use feature learning to automatically learn better representations. Some of the most successful feature learning techniques are based on unsupervised or supervised nonnegative matrix factorization (NMF) variants [11,12,13]. Finally, ASC is nowadays mainly being addressed with deep neural network-based techniques.…”

“…The usual approach to sound classification with deep learning is to count on the intermediate layers of the network to extract meaningful information for classification from time-frequency representations or even waveforms [18]. However in ASC, it has been showed that NMFbased feature learning can be competitive with deep learning techniques even when using linear classifiers [13]. Therefore, one can expect that decomposing time-frequency representa-tions with NMF could better fill the role of the first layer of the network by providing suitable features for the task.…”