10.5281/zenodo.823695
https://zenodo.org/records/823695
oai:zenodo.org:823695
Shibata Hayato
Shibata Hayato
Tokyo Institute of Technology
Kato Taku
Kato Taku
Tokyo Institute of Technology
Shinozaki Takahiro
Shinozaki Takahiro
Tokyo Institute of Technology
Watanabe Shinji
Watanabe Shinji
Mitsubishi Electric Research Laboratories
Composite Embedding Systems Based on DNN-HMM and Attention End-To-End for ZeroSpeech2017 track1 (2)
Zenodo
2017
DNN-HMM
attention end-to-end system
bottleneck
PCA
2017-07-06
10.5281/zenodo.823694
https://zenodo.org/communities/zerospeech2017
Creative Commons Attribution 4.0 International
Deep neural networks (DNNs) were trained for posterior and bottleneck features using Japanese and other language speech data. We explore various DNN types, their combinations, and dimension reduction by principal component analysis (PCA).
This version (version 2 ) concatenates CSJ feature vector and PCA compressed feature vector made from attention end-to-end feature.
X:CSJ feature (60 dim bottleneck, (version 1 feature))
S:Attention end-to-end feature (320 dim)
T:PCA(S) (60 dim)
Z=concat(X,T)