Acoustic scene classification (ASC) aims to classify a given audio recording into one of a predefined set of classes, representing the environment where the recording was made. A practical acoustic scene classification system must be robust to different sources of variation. Variations in the acoustic scene data can be caused by domain shift, which may be influenced by several factors, including unseen recording devices (recording devices used in test data are different from train data). In this paper, we utilize transformer-based features to represent acoustic scene data. To achieve device generalization, we consider different devices as different domains and use the common specific decomposition method to learn a device-independent classifier. Our experiments on a standard acoustic scene dataset containing data recorded from multiple devices, show that transformer-based features are more robust to domain shift when compared to convolutional neural network-based features. Moreover, experiments on a variety of other devices demonstrate that the use of device-independent classifier results in improved device generalization, which is a desirable attribute of a practical acoustic scene classification system.