AudioHero

AudioHero is a real-time sound-based danger detection system.

Ongoing

• Defining a context to use
• Download raw data
• Data Pre-processing
• Featurization
• Modify Network
• Training
• Evaluation
• Prepare live demo

DataSet

We use AudioSet, a dataset of over 2 million human-labeled 10-second YouTube video soundtracks. We extracted only dangerous sounds on Audioset. Please check dataset folder in this repository.

Pre-processing

1. Pre-processing
2. Amplitude Augmentation
3. Persistence Augmentation
4. Mixing Augmentation

This processing procedure referred to the paper Ubicoustics.

Result

AudioHero got total 73% classification accuracy over 8 danger situation classes.