Multi-level distance embedding learning for robust acoustic scene classification with unseen devices

Acoustic scene classification (ASC) aims to analyse the recording scene of a piece of audio. In real life, ASC has to deal with audio data from various recording devices, even those recorded by devices that did not appear during the training phase. Audio data recorded by different devices, especially unseen devices, have differences in sampling rate, amplitude, data distribution, etc. These differences can greatly interfere with the feature learning process of CNNs and lead to degradation of the performance of the ASC model. In order to learn advanced features that are less susceptible to differences in device information from manual features that contain device information, we propose an ASC method based on multi-level distance embedding space, called multi-level distance embedding learning (MDEL). There is a hierarchical relationship among the categories of acoustic scene, that is, from the three coarse-grained categories of indoor, outdoor, and transportation to more fine-grained categories. This relation corresponds to a similarity relation between categories of different granularity. MDEL exploits this hierarchical relationship of similarity between acoustic scene classes to construct embedding space containing multi-level distance. During the learning process, the model is guided to focus more on common features of the same scene classes and learn an advanced feature that is more robust to the device, thus improving the robustness of the model to data from unseen devices. Our method was evaluated on the audio dataset provided by the DCASE2020 Challenge for Task1a, and the overall classification accuracy was improved by 1.2\(\%\). For audio data from unseen devices, the classification accuracy was improved by 2.3\(\%\).