2024-03-29T00:56:32Zhttps://ipsj.ixsq.nii.ac.jp/ej/?action=repository_oaipmhoai:ipsj.ixsq.nii.ac.jp:000626662023-04-27T10:00:04Z01164:05159:05160:05697
An Investigation of Hidden Structure ModelAn Investigation of Hidden Structure Modeleng音響モデルhttp://id.nii.ac.jp/1001/00062666/Technical Reporthttps://ipsj.ixsq.nii.ac.jp/ej/?action=repository_action_common_download&item_id=62666&item_no=1&attribute_id=1&file_no=1Copyright (c) 2009 by the Information Processing Society of JapanGraduate School of Information Science and Technology, The University of TokyoGraduate School of Engineering, The University of TokyoGraduate School of Information Science and Technology, The University of TokyoYu, QIAOMasayuki, SUZUKINobuaki, MINEMATSUIn recent years, we have been working toward a structural representation of speech using contrastive features that are robust to non-linguistic variations. This paper generalizes the structural representation to Hidden Structure Model (HSM) by introducing hidden states and probabilistic calculations. HSM not only can solve miss-alignment problems of events, but also can conduct structure-based decoding, which allows us to apply HSM to general speech recognition tasks. This paper focuses on the fundamental theories of HSM. Different from HMM, HSM accounts for both the absolute and contrastive aspects of an input sequence. We show that the state inference of HSM can be formulated as a quadratical programming problem. We also introduce EM algorithm to estimate the parameters of HSM.In recent years, we have been working toward a structural representation of speech using contrastive features that are robust to non-linguistic variations. This paper generalizes the structural representation to Hidden Structure Model (HSM) by introducing hidden states and probabilistic calculations. HSM not only can solve miss-alignment problems of events, but also can conduct structure-based decoding, which allows us to apply HSM to general speech recognition tasks. This paper focuses on the fundamental theories of HSM. Different from HMM, HSM accounts for both the absolute and contrastive aspects of an input sequence. We show that the state inference of HSM can be formulated as a quadratical programming problem. We also introduce EM algorithm to estimate the parameters of HSM.AN10442647研究報告音声言語情報処理(SLP)2009-SLP-775162009-07-102009-08-19