Train and evaluation of sound-event detector.

Train a CNN detector:

Train 2d CNN on wavesound spectogram image or a 1d CNN on raw sound wave samples

run main.py

Train an SVM detector:

Train an SVM on Spectogram columns or frames of raw sound wave samples

run main.py

Requirements

soundfile
librosa

Credits

Greatly inspired by https://github.com/qiuqiangkong/dcase2019_task3

Sound event detection example

This image is an evaluation of a detector working on the spectogram domain:

Top: A spectogram of 60s sound file.
Middle: Event prediction for each frame (temporal segment of sound wave). Confidence values from 0 to 1.
Bottom: Event ground-truth for each frame. Confidence values from 0 to 1.