On the Separability of Local SNR Values Represented by Amplitude Modulation Spectrogram Vectors

DOI

10.2991/jiaet-18.2018.49How to use a DOI?

Keywords

Amplitude modulation spectrograms; Signal to noise ratio (SNR); Deep belief networks; t-distributed stochastic neighbor embedding (t-SNE).

Abstract

Evaluating local SNR values in noisy speech is very valuable for many applications. Research in this paper investigates the separability of local SNR values which are represented by amplitude modulation spectrogram (AMS) feature vectors. Deep belief networks (DBN) were employed in previous studies for binary and quarterly classifications of these SNR feature vectors, however with unsatisfactory results. It is difficult to identify which factor, AMS feature vector or DBN configuration should be responsible for the result. Therefore the technique of t-distributed stochastic neighbor embedding (t-SNE) is called in this study to visualize the separability of AMS feature vectors. According to experimental observations with binary and quarterly classifications, AMS feature vector is a good representation of SNR values with fine grain separability. To improve the classification performance of SNR values represented by AMS feature vectors, more attention should be put on DBN’s training and configuration.

Copyright

© 2018, the Authors. Published by Atlantis Press.

Open Access

This is an open access article distributed under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).

Download article (PDF)

TY  - CONF
AU  - Wei Wei
AU  - Jinhuan Wang
AU  - Jing Wang
PY  - 2018/03
DA  - 2018/03
TI  - On the Separability of Local SNR Values Represented by Amplitude Modulation Spectrogram Vectors
BT  - Proceedings of the 2018 Joint International Advanced Engineering and Technology Research Conference (JIAET 2018)
PB  - Atlantis Press
SP  - 278
EP  - 283
SN  - 2352-5401
UR  - https://doi.org/10.2991/jiaet-18.2018.49
DO  - 10.2991/jiaet-18.2018.49
ID  - Wei2018/03
ER  -