APSIPA Transactions on Signal and Information Processing > Vol 6 > Issue 1

Combining augmented statistical noise suppression and framewise speech/non-speech classification for robust voice activity detection

Yasunari Obuchi, Tokyo University of Technology, Japan, obuchiysnr@stf.teu.ac.jp
 
Suggested Citation
Yasunari Obuchi (2017), "Combining augmented statistical noise suppression and framewise speech/non-speech classification for robust voice activity detection", APSIPA Transactions on Signal and Information Processing: Vol. 6: No. 1, e7. http://dx.doi.org/10.1017/ATSIP.2017.8

Publication Date: 14 Jul 2017
© 2017 Yasunari Obuchi
 
Subjects
 
Keywords
SpeechVoice activity detectionNoise suppressionConvolutional neural networkCENSREC-1-C
 

Share

Open Access

This is published under the terms of the Creative Commons Attribution licence.

Downloaded: 1351 times

In this article:
I. INTRODUCTION 
II. AUGMENTED STATISTICAL NOISE SUPPRESSION 
III. CLASSIFICATION WITHOUT MODEL TRAINING 
IV. CLASSIFICATION WITH UNSUPERVISED AND SUPERVISED MODEL TRAINING 
V. EXPERIMENTAL RESULTS 
VI. CONCLUSIONS 

Abstract

This paper proposes a new voice activity detection (VAD) algorithm based on statistical noise suppression and framewise speech/non-speech classification. Although many VAD algorithms have been developed that are robust in noisy environments, the most successful ones are related to statistical noise suppression in some way. Accordingly, we formulate our VAD algorithm as a combination of noise suppression and subsequent framewise classification. The noise suppression part is improved by introducing the idea that any unreliable frequency component should be removed, and the decision can be made by the remaining signal. This augmentation can be realized using a few additional parameters embedded in the gain-estimation process. The framewise classification part can be either model-less or model-based. A model-less classifier has the advantage that it can be applied to any situation, even if no training data are available. In contrast, a model-based classifier (e.g., neural network-based classifier) requires training data but tends to be more accurate. The accuracy of the proposed algorithm is evaluated using the CENSREC-1-C public framework and confirmed to be superior to many existing algorithms.

DOI:10.1017/ATSIP.2017.8