ACOUSTICTOARTICULATORY INVERSION Click to edit Master subtitle style
ACOUSTIC-TO-ARTICULATORY INVERSION Click to edit Master subtitle style By: İ. Yücel Özbek
Outline: q q q q q What do we mean Acoustic-to-Articulatory Inversion (AAI)? Basic difficulties of AAI Linear regression for AAI Gaussian mixture model for AAI Kalman filter for AAI Jump Markov Linear Systems (Switching Kalman) for AAI Preliminary experimental results Conclusion Feature work plan
What do we mean Acoustic-to. Articulatory Inversion (AAI)? q We estimate articulatory trajectories from acoustic data Acoustic Features (MFCC, LPC, . . etc) Electromagnetic Articulography (EMA) Trajectories
Basic difficulties of AAI q Properties of mapping function; q Analytically unknown q Non-linear (quantal nature of speech) q One-to-many (that is, almost same acoustic features can be produced different articulatory configuration).
Basic linear regression for AAI q q Let x and y be articulatory and acoustic vector, and assuming they are jointly Gaussian. Let z be joint vector of x an y The parameters A and b can be learned from training data
Gaussian mixture model for AAI q q In addition assumption given Linear Regression, suppose z is coming from Gaussian mixture model The parameters, β A and b can be learned from training data
Kalman filter for AAI FROM STATIC TO DYNAMIC ESTIMATION q Similar to Linear Regression, we can find E( y | x) By adding time index q If we add dynamic constraint on x
Kalman filter for AAI q State space representation for Acoustic to articulatory inversion q Estimate parameter set q Estimate the state: Kalman filter or Smoother
Multiple Model Linear Dynamical Systems for AAI q q q Inspirit from GMM , we can define Dynamic system model for each mixture Estimate parameter set; Estimate the state: Switching Kalman Filter- Smoother
Piece-Wise Kalman Filter for AAI (If model boundaries are known) Click to edit the outline text format Second Outline Level Click to edit the Third outline. Outline text Level format Fourth Outline Second Outline Level Fifth Outline
Jump Markov Linear Systems (Switching Kalman) for AAI (If model boundaries are unknown) Model-2 Model-3
Jump Markov Linear Systems (Switching Kalman) for AAI q q Calculation of exact posterior distribution is infeasible due to exponentially increasing terms. The approximate solutions are required. q Generalized Pseudo Bayesian {GPB(1), GPB(2), IMM} q Sampling (MCMC) q Variational Bayesian q …
Experimental Results (Database) q We used Mocha-Timit database q Total 460 sentences q 368 sentence for training 92 sentence for testing q Female speaker is used in these experiments
Preliminary Experimental Results q RMS error for GMM , GMMSmooth, SKSmooth methods
FUTURE WORK PLAN o The
Thank you
- Slides: 16