EigenScape: A Database of Spatial Acoustic Scene Recordings

Marc Ciufo Green,Damian Murphy

doi:10.3390/app7111204

Abstract

The classification of acoustic scenes and events is an emerging area of research in the field of machine listening. Most of the research conducted so far uses spectral features extracted from monaural or stereophonic audio rather than spatial features extracted from multichannel recordings. This is partly due to the lack thus far of a substantial body of spatial recordings of acoustic scenes. This paper formally introduces EigenScape, a new database of fourth-order Ambisonic recordings of eight different acoustic scene classes. The potential applications of a spatial machine listening system are discussed before detailed information on the recording process and dataset are provided. A baseline spatial classification system using directional audio coding (DirAC) techniques is detailed and results from this classifier are presented. The classifier is shown to give good overall scene classification accuracy across the dataset, with 7 of 8 scenes being classified with an accuracy of greater than 60% with an 11% improvement in overall accuracy compared to use of Mel-frequency cepstral coefficient (MFCC) features. Further analysis of the results shows potential improvements to the classifier. It is concluded that the results validate the new database and show that spatial features can characterise acoustic scenes and as such are worthy of further investigation.

Highlights

Since machine listening became an eminent field in the early 1990s, the vast majority of research has focused on automatic speech recognition (ASR) [1] and computational solutions to the well-known
The DCASE challenges have attracted a large number of submissions designed to solve the problem of acoustic scene classification (ASC) or acoustic event detection (AED)
It can be seen that using all directional audio coding (DirAC) features to train a Gaussian mixture model (GMM) classifier gives a mean accuracy of 64% across all scene classes, whereas Mel-frequency cepstral coefficient (MFCC) features give a

Summary

Introduction

Since machine listening became an eminent field in the early 1990s, the vast majority of research has focused on automatic speech recognition (ASR) [1] and computational solutions to the well-known ‘cocktail party problem’—the “ability to listen to and follow one speaker in the presence of others” [2]. This is a mature field of study, with robust speech recognition systems featured in most modern smartphones. A typical ASC or AED system requires a feature extraction stage in order to reduce the complexity of the data to be classified

Methods

Results

Conclusion