This repository contains the introductions to the datasets and code used in our paper, titled "Low-Complexity Acoustic Scene Classification Using Parallel Attention-Convolution Network" (as shown in the section of Citation).
We conduct our experiments on the TAU Urban Acoustic Scene 2022 Mobile development dataset (TAU22) which consists of audio clips acquired by mobile devices in urban environments. You can download from here.
The required library files are placed in requirements.txt
Our environment: RTX3090 + cuda11.3 + torch1.11
You need to create a directory named reuse to save training and validation data, and in the estimate_devices_freq.ipynb will used for spectrum modulation.
student.ipynb is the main code for training and validating the datasets.
Teacher models are same with the submission to the DCASE2023, and some are pretrained and provided in teacher_models.
Please cite our paper if you find the work in our paper are useful for your research.
[1] Y. Li, J. Tan, G. Chen, J. Li, Y. Si, and Q. He, "Low-Complexity Acoustic Scene Classification Using Parallel Attention-Convolution Network," in Proc. of Interspeech, 2024.