Spontaneous speech recognition using a statistical model .of VTR-dynamics
VTR-model evaluation. ---overview
What is new
What is new (con’t)
Mathematical formulation.of the model
VTR-dynamics Illustration
Stochastic nonlinear “supersegment” model
Likelihood computation
Extended Kalman Filter
ML parameter estimation of nonlinear model: EM algorithm
EM algorithm (con’t)
Economy of model parameters
Initial experiment (A)
Large-scale experiments. ---- conditions
Results: 1241 male test utterances
Speaker variation in WER
Explain why the VTR model does the right job
Model synthesis (correct hypothesis; VTR by EKF)
Model synthesis (incorrect hypothesis; EKF)
Future work (short term)
Future work (long term)
PPT Slide
Model synthesis (correct hypothesis)
Model synthesis (incorrect; model parameters)
Experimental Setup
Initial experiment (B)
Initial experiment (C)
Email: deng@crg5.uwaterloo.ca
Home Page: sip.uwaterloo.ca
Download presentation source