Deep learning for high-level sound categorization

Patrice Guyot (1)

1 IRIT-SAMoVA - Équipe Structuration, Analyse et MOdélisation de documents Vidéo et Audio

This document presents a short description of two systems for sound classification submitted at the data challenge Mak-ing Sense of Sounds in 2018. The aim of this challenge is to classify audio files into five different classes:Nature, Hu-man, Music, Effects, Urban, which were derived from human classification. For this challenge, the organizers provided a development dataset that consists of 1500 audio files of 5 second duration divided into the five categories, each containing300 files. The following parts describe the two submitted systems.The first system, called simplemind, consists of a Convolutional Neural Network (CNN). The second system, called nevermindis based on VGG-like model adapted from.

Domaines: Informatique [cs]; Apprentissage [cs.LG]

Patrice Guyot. Deep learning for high-level sound categorization. 2018. ⟨hal-02983149⟩