Low-Complexity Acoustic Scene Classification Using Data Generation Based On Primary Ambient Extraction

Acoustic scene classification (ASC) is an important branch of machine hearing. Since ASC systems are intended to be deployed on mobile devices, how to ensure the performance under low-complexity implementation has become an attracting research problem. The state-of-the-art methods include compressing parameter precisions, reducing quantization bits, introducing sparsity constraints and so on. These methods mainly focus on the model level optimization, while explorations are rarely originated from the data level. This paper introduces a train of thoughts from data level, inspired by a stereo audio processing algorithm, namely the primary ambient extraction (PAE), which generates additional samples through audio up-mixing. The experiment results demonstrate that the proposed method exhibits better performance than a group of ASC baseline systems without data level optimization, not to mention that the proposed method is compatible with the existing model level optimization.