1 / 37

Factor Analysis of Acoustic Features for Streamed Hidden Markov Modeling

Factor Analysis of Acoustic Features for Streamed Hidden Markov Modeling. Chuan-Wei Ting Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan, Taiwan. Outline. Introduction Cepstral Factor Analysis FA Streamed Hidden Markov Model Experiments

jaclyn
Télécharger la présentation

Factor Analysis of Acoustic Features for Streamed Hidden Markov Modeling

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Factor Analysis of Acoustic Features for Streamed Hidden Markov Modeling Chuan-Wei Ting Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan, Taiwan

  2. Outline • Introduction • Cepstral Factor Analysis • FA Streamed Hidden Markov Model • Experiments • Conclusions & Future Works

  3. Outline • Introduction • Stochastic modeling • Cepstral Factor Analysis • FA Streamed Hidden Markov Model • Experiments • Conclusions & Future Works

  4. Introduction • The objective of constructing acoustic model is to capture the characteristics of speech signal. • Stochastic modeling • Hidden Markov model (HMM) • Multi-Stream HMM • Factorial HMM

  5. Hidden Markov Model • Topology of HMM • Constraints • All features are “tied” together • Topology • Transition moment • Independent assumption

  6. Multi-Stream HMM • Topology of Multi-stream HMM

  7. Simplification of Multi-Stream HMM • Streams are assumed to be statistical independent • Weighted log-likelihood approach

  8. Factorial HMM • Topology of FHMM

  9. Outline • Introduction • Cepstral Factor Analysis • Features analysis • Factor analysis • FA Streamed Hidden Markov Model • Experiments • Conclusions & Future Works

  10. Cepstral Factor Analysis • Feature analysis • Dynamics of different features • Correlations

  11. Factor Analysis • Discover the correlations inherent in observation data. • Applications • Data compression • Signal processing • Acoustic modeling

  12. specific factor factor loading matrix common factor Mathematical Definition of FA • FA conducts data analysis of the multivariate observations using the common factors and the specific factors. • For a dimensional feature vector , the general form of FA model is given by

  13. Principal Component Solution • Find an estimator that will approximate the fundamental expression • Decompose covariance matrix of observation • FA parameters can be estimated by

  14. Principal Factor Analysis Solution • Using an initial estimate (diagonal) and then obtain loading matrix by • Obtain an estimate of by performing a principal component analysis on . • This process is continued until the communality estimates converge.

  15. Maximum Likelihood Solution • When FA is carried out on the correlation matrix • Where , , , , and is a diagonal matrix.

  16. Varimax rotation • Let • can be obtained by maximizing Rotation of Loading Matrix • Rotate loading matrix by an orthogonal matrix • Where satisfies

  17. Effectiveness of Rotation • Obtain greater discriminability

  18. Outline • Introduction • Cepstral Factor Analysis • FA Streamed Hidden Markov Model • Survey of different HMMs • FASHMM • Experiments • Conclusions & Future Works

  19. FA Streamed HMM • Using FA, the processes of observed features and hidden states are represented by common factors and residual factors.

  20. Survey of Different HMMs (FAHMM) • Covariance matrix modeling • Full vs. diagonal • Sufficient data problem • FA representation • State/latent representation • Discrete vs. continuous

  21. Survey of Different HMMs (Streamed HMM) • In standard HMM, the joint probability of observation sequence and state sequence was represented by • Using FHMM, the state at time was extended to states, i.e. . • Likelihood combination • Multi-stream HMM • FHMM  sub-word level  frame level

  22. common covariance matrix Likelihood Function of FHMM • State transition probability • Likelihood function

  23. Estimation Approaches for FHMM • Exact inference • Expectation maximization (EM) algorithm • Complexity • Approximations • Gibbs sampling • Variational inference

  24. FASHMM • According to FA method, the common factor are associated with some features, which are highly correlated. • Correlated features are grouped together in a stream and shared by the same FA parameters. • Observed feature vector can be represented by

  25. Topology of FASHMM • State transition probability

  26. Outline • Introduction • Cepstral Factor Analysis • FA Streamed Hidden Markov Model • Experiments • Simulated data setup • HMM vs. FASHMM • Recognition results & discussion • Conclusions & Future Works

  27. Experimental Setup • Simulated data • 4 classes, 5 variables • Training: 100 sentences, 5 “words” per sentence • Testing: 50 utterances, 4 “words” per sentence • Model structure • HMM • 7 states each class • Only one Gaussian each state • FASHMM • 3 states each class • Only one Gaussian each state

  28. Class 1

  29. Class 2

  30. Class 3

  31. Class 4

  32. HMM vs. FASHMM HMM FASHMM

  33. Recognition Results

  34. Discussion

  35. Outline • Introduction • Cepstral Factor Analysis • FA Streamed Hidden Markov Model • Experiments • Conclusions & Future Works

  36. Conclusions • We have presented the FA approach • Extract the common factor and the residual factors in acoustic features • Separate the Markov chains for these factors. • Represent the sophisticated dynamics in stochastic process of speech signal. • A new topology of FA streamed HMM was proposed.

  37. Future Works • More acoustic features • Model selection • Streams • States • Mixtures • Large vocabulary continuous speech recognition (LVCSR) task

More Related