Audio Scene Classification with Deep Recurrent Neural Networks

Abstract

We introduce in this work an efficient approach for audio scene classification using deep recurrent neural networks. An audio scene is firstly transformed into a sequence of high-level label tree embedding feature vectors. The vector sequence is then divided into multiple subsequences on which a deep GRUbased recurrent neural network is trained for sequence-to-label classification. The global predicted label for the entire sequence is finally obtained via aggregation of subsequence classification outputs. We will show that our approach obtains an F1-score of 97.7% on the LITIS Rouen dataset, which is the largest dataset publicly available for the task. Compared to the best previously reported result on the dataset, our approach is able to reduce the relative classification error by 35.3%.

Spatio-Temporal Attention Pooling for Audio Scene Classification

“…Compared to the standard softmax, Support Vector Machines (SVM) usually achieve better generalization due to their maximum margin property [26]. Similar to [6,2], after training the network, we calibrate the final classifier by employing a linear SVM in replacement for the softmax layer. The trained network is used to extract feature vectors for the original training examples (without data augmentation) which are used to train the SVM classifier.…”

“…Note that the classification label of a 30-second recording was derived via aggregation of the classification results of its 2-second segments. To this end, probabilistic multiplicative fusion, followed by likelihood maximization were carried out similar to [2].…”

“…Combining the classification results of the Att-CRNN on both types of feature using the probabilistic multiplicative aggregation [2] (i.e. Att-CRNN (fusion) in Table 3) further enlarges the margin up to 0.66% absolute gain on the overall accuracy, or…”

“…Convolutional neural network (CNNs) [6,7,8,5,4,9,10,11] are most commonly used deep learning techniques, thanks to their feature learning capability. Leveraging the nature of audio signals, sequential modelling with recurrent neural networks (RNNs) [2,12,13] and temporal transformer networks [10] have also demonstrated results on par with the convolutional counterparts. The deep learning models were trained either on time-frequency representations, such as log Mel-scale spectrograms [4,11], or highlevel features like label tree embedding features [6,2].…”

Acoustic features fusion using attentive multi-channel deep architecture

“…A popular approach for acoustic scene recognition (ASR) and the tagging task is to use the low-level or high-level acoustic features such as Mel-frequency cepstral coefficients (MFCCs), Mel-spectrogram, Mel-bank, log Mel-bank features, etc., with the state-of-the-art deep models [6,7]. Some of these acoustic features possess complementary qualities, that is, for two given features, one is apt in identifying certain specific classes, while the other is suitable for the rest.…”