A Bayesian Attention Neural Network Layer for Speaker Recognition

Neural network based attention modeling has found utility in areas such as visual analysis, speech recognition and more recently speaker recognition. Attention represents a gating (or weighting) function on information and governs how the corresponding statistics are accumulated. In the context of s...

Full description

Saved in:
Bibliographic Details
Published inICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) pp. 6241 - 6245
Main Authors Zhu, Weizhong, Pelecanos, Jason
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.05.2019
Subjects
attention modeling
Bayes methods
Bayesian statistics
deep neural networks
Neural networks
NIST
Speaker recognition
Training
Training data
Online AccessGet full text

Cover

More Information
Summary:Neural network based attention modeling has found utility in areas such as visual analysis, speech recognition and more recently speaker recognition. Attention represents a gating (or weighting) function on information and governs how the corresponding statistics are accumulated. In the context of speaker recognition, attention can be incorporated as a frame weighted mean of an information stream. These weights can be made to sum to one (the standard approach) or be calculated in other ways. If the weights can be made to represent event observation probabilities, we can extend the approach to be within a Bayesian framework. More specifically, we combine prior information with the frame weighted statistics to produce an adapted or posterior estimate of the mean. We evaluate the proposed method on NIST data.
ISSN:2379-190X
DOI:10.1109/ICASSP.2019.8682953