The repo includes code that is used in DCASE challenge of acoustic anomaly detection.
- It ranks 16/40 in the challenge.
- It is the a sumbission withou using any machine learning.
- Each sound is represented by an Gaussian distribution of MFCCs
- The dissimilarity of each sound pair is measured by KL divergence
- Anomaly score is determined by the most similar normal sound of the same category
The experiment data and results of submissions are available https://dcase.community/challenge2020/task-unsupervised-detection-of-anomalous-sounds
- Download the data.
- Revise the data root in param.yaml
- python compute_mfcc_dataset.py
- python predict_kl.py (or predict_kl2.py, or predict_kl3.py)
- python evaluate.py