pipeline_tag: audio-classification
This repository contains the models submitted to Task 1 of the DCASE 2024 Challenge
Description
The task is to develop a data-efficient and low-complexity acoustic scene classification system.
The challenge dataset consists of 1 second audio clips from one of 10 classes: airport, bus, metro, metro_station, park, public_square, shopping_mall, street_pedestrian, street_traffic, tram. Five models are trained on splits of the training data: 5%, 10%, 25%, 50%, and 100%, respectively.
We chose to use the baseline model architecture and apply a target-specific training process which involves a pretraining dataset that is pruned to match the target dataset. Knowledge distillation is used to transfer knowledge from a pre-trained audio tagging ensemble to the target model. A technical report describing the training process can be found here
Results
The full results of all participants can be found here: https://dcase.community/challenge2024/task-data-efficient-low-complexity-acoustic-scene-classification-results
The results of our submission compared to the baseline on the evaluation data are as follows:
| Name | Official rank | Rank score | Split 5% | Split 10% | Split 25% | Split 50% | Split 100% |
|---|---|---|---|---|---|---|---|
| Werning_UPBNT | 8 | 54.35 | 49.21 % | 52.51 % | 55.49 % | 56.20 % | 58.34 % |
| Baseline | 17 | 50.73 | 44.00 % | 46.95 % | 51.47 % | 54.40 % | 56.84 % |
Usage
The example notebook shows how to predict the acoustic scene for a given audio file using the models.
The model code is adapted from the baseline repository: https://github.com/CPJKU/dcase2024_task1_baseline