Distilling the Knowledge of Specialist DNNs in Acoustic Scene Classification

This repository contains script and DNN models that was used for the DCASE2019 challenge task1-a. Currently, there are only codes for raw waveform model. Overall description of the system is in the Workshop paper and implementation details are further dealt in the Technical report.
(for now, the Workshop paper link is connected to our previous work on Knowledge distillation in acoustic scene classification, which will be presented at Interspeech 2019)

Introduction

Common acoustic properties among different acoustic scenes were pointed as one of the causes for performance degradation in acoustic scene classification (ASC) task. ¹ These common properties resulted in a few pairs of acoustic scenes that are frequently misclassified (see the left confusion matrix in below image). In our Workshop paper ², we use the concept of specialist models that is in Hinton et al.'s paper ³, modifying for ASC.

Specialist Knowledge Distillation

How to use scripts

Reference

1: H. Heo, J. Jung, H. Shim and H. Yu, Acoustic scene classification using teacher-student learning with soft-labels, Interspeech 2019 2: J. Jung, H. Heo, H. Shim and H. Yu, DISTILLING THE KNOWLEDGE OF SPECIALIST DEEP NEURAL NETWORKS IN ACOUSTIC SCENE CLASSIFICATION, DCASE 2019 Workshop
3: G. Hinton, O. Vinyals, and J. Dean, Distilling the Knowledge in a Neural Network, NIPS 2014 deep learning workshop