Search

Deep convolutional neural networks (DCNNs) has been widely used in music auto-tagging which is a multi-label classification task that predicts tags of audio signals. This paper presents a sample-level DCNN for music auto-tagging. The proposed DCNN highlights two components: strided convolutional layer for extracting local feature and reducing temporal dimension, and residual block from WaveNet for preserving input resolution and extracting more complex features. In order to further improve performance, squeeze-and-excitation (SE) block is introduced to the residual block. Under the evaluation metric of Area Under Receiver Operating Characteristic Curve (AUC-ROC) score, experiment results on MagnaTagATune (MTAT) dataset show that the two proposed models achieve 91.47% and 92.76% respectively. Furthermore, our proposed models have slightly surpass the state-of-the-art model SampleCNN with SE block.

In this paper, an end-to-end speaker verification system based on attentive deep convolutional neural network (CNN) is highlighted. It takes log filter bank coefficients as input and measures speaker similarity between a test utterance and enrollment utterances by cosine similarity for verification. The approach utilizes the channel attention module of convolutional block attention module (CBAM) to increase representation power by giving different weights to feature maps. In addition, softmax is used to pre-train for initializing the weights of the network and tuple-based end-to-end (TE2E) loss function is responsible for fine-tune in evaluation stage, such a strategy not only results in notable improvements over the baseline model but also allows for direct optimization of the evaluation metric. Experimental results on VoxCeleb dataset indicates that proposed model achieves an equal error rate (EER) of 3.83%, which is slightly worse than x-vectors while outperforms i-vectors.

This paper highlights a structure called SECNN. It combines squeeze-and-excitation (SE) components with a simplified residual convolutional neural network (ResNet). This model takes time-frequency spectrogram as input and measures speaker similarity between an utterance embedding and speaker models by cosine similarity. Speaker models are obtained by averaging utterance level features of each enrollment speaker. On the one hand, SECNN can mitigate speaker overfitting in speaker verification by using some techniques such as regularization techniques and SE operation. On the other hand, SECNN is a lightweight model with merely 1.5M parameters. Experimental results indicate that SECNN outperforms other end-to-end models such as Deep Speaker and achieves an equal error rate (EER) of 5.55% in speaker verification and accuracy of 93.92% in speaker identification on Librispeech dataset. It also achieves an EER of 2.58% in speaker verification and accuracy of 95.83% in speaker identification on TIMIT dataset.

In recent years, convolutional neural networks (CNNs) become popular approaches used in music information retrieval (MIR) tasks, such as mood recognition, music auto-tagging and so on. Since CNNs are able to extract the local features effectively, previous attempts show great performance on music auto-tagging. However, CNNs is not able to capture the spatial features and the relationship between low-level features are neglected. Motivated by this problem, a hybrid architecture is proposed based on Capsule Network, which is capable to extract spatial features with the routing-by-agreement mechanism. The proposed model was applied in music auto-tagging. The results show that it achieves promising results of the ROC-AUC score of 90.67%.