A Non-linear GMM KL and GUMI Kernel for SVM Using GMM-UBM Supervector in Home Acoustic Event Classification

Ngoc Nam BUI  Jin Young KIM  Tan Dat TRINH  

IEICE TRANSACTIONS on Fundamentals of Electronics, Communications and Computer Sciences   Vol.E97-A   No.8   pp.1791-1794
Publication Date: 2014/08/01
Online ISSN: 1745-1337
DOI: 10.1587/transfun.E97.A.1791
Type of Manuscript: LETTER
Category: Digital Signal Processing
non-linear GMM KL,  non-linear GUMI,  audio event recognition,  GMM supervector,  kernel combination,  

Full Text: PDF(630.7KB)>>
Buy this Article

Acoustic Event Classification (AEC) poses difficult technical challenges as a result of the complexity in capturing and processing sound data. Of the various applicable approaches, Support Vector Machine (SVM) with Gaussian Mixture Model (GMM) supervectors has been proven to obtain better solutions for such problems. In this paper, based on the multiple kernel selection model, we introduce two non-linear kernels, which are derived from the linear kernels of GMM Kullback-Leibler divergence (GMM KL) and GMM-UBM mean interval (GUMI). The proposed method improved the AEC model's accuracy from 85.58% to 90.94% within the domain of home AEC.