講演抄録/キーワード |
講演名 |
2016-03-28 13:15
[ポスター講演]An evaluation of acoustic-to-articulatory inversion mapping with latent trajectory Gaussian mixture model ○Patrick Lumban Tobing(NAIST)・Tomoki Toda(Nagoya Univ./NAIST)・Hirokazu Kameoka(NTT)・Satoshi Nakamura(NAIST) EA2015-85 SIP2015-134 SP2015-113 |
抄録 |
(和) |
In this report, we present an evaluation of acoustic-to-articulatory inversion mapping based on latent trajectory
Gaussian mixture model (LTGMM). In a conventional GMM-based inversion mapping system, GMM parameters
are optimized by maximizing the likelihood of joint static and dynamic features of acoustic-articulatory data.
In the mapping process, given the acoustic data, smoothly varying
articulatory parameter trajectories are estimated by maximizing the
conditional likelihood of their static features only, where the
inter-frame correlation is taken into account by imposing the explicit
relationship between static and dynamic features. Because training and optimization criteria are different from each other,
the trained GMM is not optimum for the mapping process. A trajectory training method has been proposed to address this inconsistency problem [1]. However, this method has difficulties in optimization of some parameters,
such as covariance matrices and a mixture component sequence. In this report, as another method to address the inconsistency problem,
we propose an inversion mapping method based on latent trajectory GMM,
inspired by the latent trjectory hidden Markov model [2]. The proposed
method makes it possible to apply EM algorithm to model parameter
optimization, which is difficult in the conventional trajectory training
method. The experimental results demonstrate that the proposed LTGMM method
outperforms the conventional GMM for the acoustic-to-articulatory inversion mapping task with lower values
of root-mean-square error and higher values of correlation coefficient. |
(英) |
In this report, we present an evaluation of acoustic-to-articulatory inversion mapping based on latent trajectory
Gaussian mixture model (LTGMM). In a conventional GMM-based inversion mapping system, GMM parameters
are optimized by maximizing the likelihood of joint static and dynamic features of acoustic-articulatory data.
In the mapping process, given the acoustic data, smoothly varying
articulatory parameter trajectories are estimated by maximizing the
conditional likelihood of their static features only, where the
inter-frame correlation is taken into account by imposing the explicit
relationship between static and dynamic features. Because training and optimization criteria are different from each other,
the trained GMM is not optimum for the mapping process. A trajectory training method has been proposed to address this inconsistency problem [1]. However, this method has difficulties in optimization of some parameters,
such as covariance matrices and a mixture component sequence. In this report, as another method to address the inconsistency problem,
we propose an inversion mapping method based on latent trajectory GMM,
inspired by the latent trjectory hidden Markov model [2]. The proposed
method makes it possible to apply EM algorithm to model parameter
optimization, which is difficult in the conventional trajectory training
method. The experimental results demonstrate that the proposed LTGMM method
outperforms the conventional GMM for the acoustic-to-articulatory inversion mapping task with lower values
of root-mean-square error and higher values of correlation coefficient. |
キーワード |
(和) |
acoustic-to-articulatory inversion mapping / Gaussian mixture model / trajectory training / inter-frame correlation / EM algorithm / / / |
(英) |
acoustic-to-articulatory inversion mapping / Gaussian mixture model / trajectory training / inter-frame correlation / EM algorithm / / / |
文献情報 |
信学技報, vol. 115, no. 523, SP2015-113, pp. 111-116, 2016年3月. |
資料番号 |
SP2015-113 |
発行日 |
2016-03-21 (EA, SIP, SP) |
ISSN |
Print edition: ISSN 0913-5685 Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
EA2015-85 SIP2015-134 SP2015-113 |
|