Search

In this letter, we propose a concise feature representation framework for acoustic scene classification by pruning embeddings obtained from SoundNet, a deep convolutional neural network. We demonstrate that the feature maps generated at various layers of SoundNet have redundancy. The proposed singular value decomposition based method reduces the redundancy while relying on the assumption that useful feature maps produced by different classes lie along different directions. This leads to ignoring the feature maps that produce similar embeddings for different classes. In the context of using an ensemble of classifiers on the various layers of SoundNet, pruning the redundant feature maps leads to reduction in dimensionality and computational complexity. Our experiments on acoustic scene classification demonstrate that ignoring 73% of feature maps reduces the performance by less than 1% with 12.67% reduction in computational complexity. In addition to this, we also show that the proposed pruning framework can be utilized to remove filters in the SoundNet network architecture, with 13x lesser model storage requirement. Also, the number of parameters reduce from 28 million to 2 million with marginal degradation in performance. This small model obtained after applying the proposed pruning procedure is evaluated on different acoustic scene classification datasets, and shows excellent generalization ability.

State-of-the-art spoken language identification (LID) networks are vulnerable to channel-mismatch that occurs due to the differences in the channels used to obtain the training and testing samples. The effect of channel-mismatch is severe when the training dataset contains very limited channel diversity. One way to address channel-mismatch is by learning a channel-invariant representation of the speech using adversarial multi-task learning (AMTL). But, AMTL approach cannot be used when the training samples do not contain the corresponding channel labels. To address this, we propose an auxiliary within-sample similarity loss (WSSL) which encourages the network to suppress the channel-specific contents in the speech. This does not require any channel labels. Specifically, WSSL gives the similarity between a pair of embeddings of same sample obtained by two separate embedding extractors. These embedding extractors are designed to capture similar information about the channel, but dissimilar LID-specific information in the speech. Furthermore, the proposed WSSL improves the noise-robustness of the LID-network by suppressing the background noise in the speech to some extent. We demonstrate the effectiveness of the proposed approach in both seen and unseen channel conditions using a set of datasets having significant channel-mismatch.

We introduce a new classification framework that combines the characteristics of matrix factorization with the discriminative capabilities of kernel methods. Short-time analysis of audio signals having different durations result in sets of feature vectors having different cardinalities. Support vector machines handle such varying-length feature sets using dynamic kernels, such as the intermediate matching kernel (IMK). IMK works by utilizing the so-called virtual vectors which select pairs of feature vectors to learn discrimination between classes. Existing formulations of IMK choose virtual vectors from the most information-bearing regions of classes, such as cluster means. This form of IMK completely ignores the feature vector pairs that lie around the class boundaries. To overcome this limitation, we propose an alternative formulation of IMK based on archetypal analysis (AA) and deep archetypal analysis (DAA). AA represents the data in terms of boundary elements, whereas DAA represents data in terms of both boundary and average elements. The proposed AA and DAA based intermediate matching kernel (AA/DAA-IMK) utilizes the elements generated from AA and DAA as the virtual feature vectors. Experimental evaluation on four different bioacoustic datasets show that the introduction of AA and DAA into the IMK framework leads to a noticeable improvement in classification accuracy.

State-of-the-art systems for spoken language identification (LID) use i-vector or embedding extracted using a deep neural network (DNN) to represent the utterance. These fixed-length representations are obtained without explicitly considering the relevance of individual frame-level feature vectors in deciding the class label. In this paper, we propose a new method to represent the utterance that considers the relevance of the individual frame-level features. The proposed representation can also preserve the locally available LID-specific information in the input features to some extent. To better utilize the local-level information in the new representation, we propose a novel segment-level matching kernel based support vector machine (SVM) classifier. The proposed representation of the utterance based on the relevance of frame-level features improves the robustness of the LID system to different background noise conditions in the speech. The experiments conducted on speech with different background conditions show that the proposed approach performs better than state-of-the-art approaches in noisy speech and performs similarly to the state-of-the-art systems in clean speech condition.

Bioacoustic classification often suffers from the lack of labeled data. This hinders the effective utilization of state-of-the-art deep learning models in bioacoustics. To overcome this problem, the authors propose a deep metric learning-based framework that provides effective classification, even when only a small number of per-class training examples are available. The proposed framework utilizes a multiscale convolutional neural network and the proposed dynamic variant of the triplet loss to learn a transformation space where intra-class separation is minimized and inter-class separation is maximized by a dynamically increasing margin. The process of learning this transformation is known as deep metric learning. The triplet loss analyzes three examples (referred to as a triplet) at a time to perform deep metric learning. The number of possible triplets increases cubically with the dataset size, making triplet loss more suitable than the cross-entropy loss in data-scarce conditions. Experiments on three different publicly available datasets show that the proposed framework performs better than existing bioacoustic classification methods. Experimental results also demonstrate the superiority of dynamic triplet loss over cross-entropy loss in data-scarce conditions. Furthermore, unlike existing bioacoustic classification methods, the proposed framework has been extended to provide open-set classification.