论文部分内容阅读
Acoustic Scene Classification (ASC) is the recognition and categorization of audio data that identifies the environment which it has been recorded. ASC is quite a challenging application of machine listening due to the noisy nature of the audio signal.
We analyze several state-of-the-art models for ASC on two different datasets. The datasets belong to the IEEE challenge on the Detection and Classification of Acoustic Scenes and Events (DCASE) 2017 and 2019. The two datasets are publicly available for research purpose, and consist of 10 and 15 classes of acoustic scenes respectively. In total, 57 hours of stereo recordings are available which includes common indoor and outdoor environmental scenes, such as beach, city center, library, forest path train, car, etc.
We propose a method of ASC via fusing the voting of deep neural networks. In the proposed method, two different acoustic features are first extracted from each audio recording i.e. Mel-frequency cepstral coefficients (MFCC) and Logarithmic filter-bank (LFB). These features are then fed into three different classifiers (deep neural networks):Visual Geometry Group (VGG), Residual Network (ResNet) and Long Short-Term Memory (LSTM). The motivation for choosing these variety of neural networks is that they have complementary advantages for ASC. After training each network and acquiring the results, fusion of classifiers voting is used to determine a final outcome. The fusion of all the results through voting technique is one of the methods under ensemble learning. The final Classification Accuracies (CA) that are obtained after the fusion of the classifiers are 73.27% and 76.99% on DCASE 2017 and 2019 datasets respectively. The proposed fusion of classifiers voting obtains the CA improvements by 12.71% and 13.79% on DCASE 2017 and 2019 datasets respectively, compared to the individual baseline classifiers.
We analyze several state-of-the-art models for ASC on two different datasets. The datasets belong to the IEEE challenge on the Detection and Classification of Acoustic Scenes and Events (DCASE) 2017 and 2019. The two datasets are publicly available for research purpose, and consist of 10 and 15 classes of acoustic scenes respectively. In total, 57 hours of stereo recordings are available which includes common indoor and outdoor environmental scenes, such as beach, city center, library, forest path train, car, etc.
We propose a method of ASC via fusing the voting of deep neural networks. In the proposed method, two different acoustic features are first extracted from each audio recording i.e. Mel-frequency cepstral coefficients (MFCC) and Logarithmic filter-bank (LFB). These features are then fed into three different classifiers (deep neural networks):Visual Geometry Group (VGG), Residual Network (ResNet) and Long Short-Term Memory (LSTM). The motivation for choosing these variety of neural networks is that they have complementary advantages for ASC. After training each network and acquiring the results, fusion of classifiers voting is used to determine a final outcome. The fusion of all the results through voting technique is one of the methods under ensemble learning. The final Classification Accuracies (CA) that are obtained after the fusion of the classifiers are 73.27% and 76.99% on DCASE 2017 and 2019 datasets respectively. The proposed fusion of classifiers voting obtains the CA improvements by 12.71% and 13.79% on DCASE 2017 and 2019 datasets respectively, compared to the individual baseline classifiers.