Presentation | 2012-07-21 WFST-based Structured Classification of Features Extracted by Using Deep Neural Networks Yotaro KUBO, Takaaki HORI, Atsushi NAKAMURA, |
---|---|
PDF Download Page | PDF download Page Link |
Abstract(in Japanese) | (See Japanese page) |
Abstract(in English) | Multilayer perceptions, which include more than 2 hidden layers, are known to be efficient for modeling of complex classification processes. However, due to the local optima and plateaus in their training objective functions, these perceptrons had not been used in practice. Recently, a heuristic method that involves the use of initial value obtained by applying unsupervised training of neural networks have enabled the practical use of such perceptrons. By introducing multiple hidden layers, the total number of needed units to accurately model the nonlinear classification processes would become smaller than that in single hidden layer networks. Consequently, we can analyze that the main contribution of introducing deep processings is enhancement in feature representations. On the other hand, an approach called structured classification have been collecting attention of speech researchers since it realizes direct modeling of sequence-to-sequence classification. However, it is known that the feature transformation is important in this approach since it typically considers the sequence classification as linear classification processes. In this paper, we attempt to combine these two approaches in order to enhance the both sides; feature representations and label representations. Specifically, we introduced the structured classification method based on weighted finite-state transducers into the multilayer perceptron-based speech recognition systems. |
Keyword(in Japanese) | (See Japanese page) |
Keyword(in English) | Automatic Speech Recognition / Structured Classification / Deep Learning / Temporal Features |
Paper # | SP2012-57 |
Date of Issue |
Conference Information | |
Committee | SP |
---|---|
Conference Date | 2012/7/12(1days) |
Place (in Japanese) | (See Japanese page) |
Place (in English) | |
Topics (in Japanese) | (See Japanese page) |
Topics (in English) | |
Chair | |
Vice Chair | |
Secretary | |
Assistant |
Paper Information | |
Registration To | Speech (SP) |
---|---|
Language | JPN |
Title (in Japanese) | (See Japanese page) |
Sub Title (in Japanese) | (See Japanese page) |
Title (in English) | WFST-based Structured Classification of Features Extracted by Using Deep Neural Networks |
Sub Title (in English) | |
Keyword(1) | Automatic Speech Recognition |
Keyword(2) | Structured Classification |
Keyword(3) | Deep Learning |
Keyword(4) | Temporal Features |
1st Author's Name | Yotaro KUBO |
1st Author's Affiliation | NTT Communication Science Laboratories() |
2nd Author's Name | Takaaki HORI |
2nd Author's Affiliation | NTT Communication Science Laboratories |
3rd Author's Name | Atsushi NAKAMURA |
3rd Author's Affiliation | NTT Communication Science Laboratories |
Date | 2012-07-21 |
Paper # | SP2012-57 |
Volume (vol) | vol.112 |
Number (no) | 141 |
Page | pp.pp.- |
#Pages | 6 |
Date of Issue |