Your browser does not support the video tag. Please use IE9+ or Google Chrome.
投影片 1 (Lab531, 1:10:09)
 
 
 
  • 1. Slide 1
  • 2. Minimum-Classification-Error (MCE) and Discriminative Training
  • 3. Slide 1
  • 4. Minimum-Classification-Error (MCE) and Discriminative Training
  • 5. Minimum-Classification-Error (MCE)
  • 6. Minimum-Classification-Error (MCE) and Discriminative Training
  • 7. Minimum-Classification-Error (MCE)
  • 8. Minimum-Classification-Error (MCE) and Discriminative Training
  • 9. Minimum-Classification-Error (MCE)
  • 10. Minimum-Classification-Error (MCE) Training
  • 11. Minimum-Classification-Error (MCE)
  • 12. Minimum-Classification-Error (MCE) and Discriminative Training
  • 13. Minimum-Classification-Error (MCE)
  • 14. Minimum-Classification-Error (MCE) Training
  • 15. Minimum-Classification-Error (MCE)
  • 16. Minimum-Classification-Error (MCE) Training
  • 17. 1 𝑑 = 1 1+𝑒𝑥𝑝 −𝛾(𝑑−θ) l(d) →0 when d →-∞l(d) →1 when d →∞θ: switching from 0 to 1 near θγ: determining the slope at switching point
  • 18. Minimum-Classification-Error (MCE) Training
  • 19. 1 𝑑 = 1 1+𝑒𝑥𝑝 −𝛾(𝑑−θ) l(d) →0 when d →-∞l(d) →1 when d →∞θ: switching from 0 to 1 near θγ: determining the slope at switching point
  • 20. Minimum-Classification-Error (MCE) Training
  • 21. Minimum-Classification-Error (MCE)
  • 22. Minimum-Classification-Error (MCE) Training
  • 23. 1 𝑑 = 1 1+𝑒𝑥𝑝 −𝛾(𝑑−θ) l(d) →0 when d →-∞l(d) →1 when d →∞θ: switching from 0 to 1 near θγ: determining the slope at switching point
  • 24. Minimum-Classification-Error (MCE) Training
  • 25. 1 𝑑 = 1 1+𝑒𝑥𝑝 −𝛾(𝑑−θ) l(d) →0 when d →-∞l(d) →1 when d →∞θ: switching from 0 to 1 near θγ: determining the slope at switching point
  • 26. Minimum-Classification-Error (MCE) Training
  • 27. 1 𝑑 = 1 1+𝑒𝑥𝑝 −𝛾(𝑑−θ) l(d) →0 when d →-∞l(d) →1 when d →∞θ: switching from 0 to 1 near θγ: determining the slope at switching point
  • 28. Minimum-Classification-Error (MCE) Training
  • 29. 1 𝑑 = 1 1+𝑒𝑥𝑝 −𝛾(𝑑−θ) l(d) →0 when d →-∞l(d) →1 when d →∞θ: switching from 0 to 1 near θγ: determining the slope at switching point
  • 30. Minimum-Classification-Error (MCE) Training
  • 31. Gradient Descent Algorithm
  • 32. Minimum-Classification-Error (MCE) Training
  • 33. Gradient Descent Algorithm
  • 34. Minimum-Classification-Error (MCE) Training
  • 35. Gradient Descent Algorithm
  • 36. Minimum-Classification-Error (MCE) Training
  • 37. 1 𝑑 = 1 1+𝑒𝑥𝑝 −𝛾(𝑑−θ) l(d) →0 when d →-∞l(d) →1 when d →∞θ: switching from 0 to 1 near θγ: determining the slope at switching point
  • 38. Minimum-Classification-Error (MCE) Training
  • 39. Minimum-Classification-Error (MCE)
  • 40. Minimum-Classification-Error (MCE) Training
  • 41. 1 𝑑 = 1 1+𝑒𝑥𝑝 −𝛾(𝑑−θ) l(d) →0 when d →-∞l(d) →1 when d →∞θ: switching from 0 to 1 near θγ: determining the slope at switching point
  • 42. Minimum-Classification-Error (MCE) Training
  • 43. 1 𝑑 = 1 1+𝑒𝑥𝑝 −𝛾(𝑑−θ) l(d) →0 when d →-∞l(d) →1 when d →∞θ: switching from 0 to 1 near θγ: determining the slope at switching point
  • 44. Minimum-Classification-Error (MCE) Training
  • 45. Minimum-Classification-Error (MCE)
  • 46. Minimum-Classification-Error (MCE) Training
  • 47. 1 𝑑 = 1 1+𝑒𝑥𝑝 −𝛾(𝑑−θ) l(d) →0 when d →-∞l(d) →1 when d →∞θ: switching from 0 to 1 near θγ: determining the slope at switching point
  • 48. Minimum-Classification-Error (MCE) Training
  • 49. Gradient Descent Algorithm
  • 50. Discriminative Training and Minimum Phone Error Rate (MPE) Training For Large Vocabulary Speech Recognition
  • 51. LatticePhone Accuracy
  • 52. Discriminative Training and Minimum Phone Error Rate (MPE) Training For Large Vocabulary Speech Recognition
  • 53. LatticePhone Accuracy
  • 54. Discriminative Training and Minimum Phone Error Rate (MPE) Training For Large Vocabulary Speech Recognition
  • 55. Gradient Descent Algorithm
  • 56. Discriminative Training and Minimum Phone Error Rate (MPE) Training For Large Vocabulary Speech Recognition
  • 57. LatticePhone Accuracy
  • 58. Discriminative Training and Minimum Phone Error Rate (MPE) Training For Large Vocabulary Speech Recognition
  • 59. LatticePhone Accuracy
  • 60. Discriminative Training and Minimum Phone Error Rate (MPE) Training For Large Vocabulary Speech Recognition
  • 61. LatticePhone Accuracy
  • 62. “ Minimum Classification Error Rate Methods for Speech Recognition”, IEEE Trans. Speech and Audio Processing, May 1997“Segmental Minimum Bayes-Rick Decoding for Automatic Speech Recognition”, IEEE Trans. Speech and Audio Processing, 2004“Minimum Phone Error and I-smoothing for Improved Discriminative Training”, International Conference on Acoustics, Speech and Signal Processing, 2002“Discriminative Training for Automatic Speech Recognition”, IEEE Signal Processing Magazine, Nov 2012
  • 63. Subspace Gaussian Mixture Model
  • 64. Subspace Gaussian Mixture Model
  • 65. Subspace Gaussian Mixture Model
  • 66. Subspace Gaussian Mixture Model
  • 67. Subspace Gaussian Mixture Model
  • 68. Subspace Gaussian Mixture Model
  • 69. Subspace Gaussian Mixture Model
  • 70. Subspace Gaussian Mixture Model
  • 71. Subspace Gaussian Mixture Model
  • 72. Subspace Gaussian Mixture Model
  • 73. Subspace Gaussian Mixture Model
  • 74. References for Subspace Gaussian Mixture Model
  • 75. Neural Network — Classification Task
  • 76. Neural Network — 2D Feature Space
  • 77. Neural Network ‒ Multi-Dimensional Feature Space
  • 78. Neural Network — Neurons
  • 79. Neural Network
  • 80. Neural Network Training – Back Propagation
  • 81. Gradient Descent Algorithm
  • 82. Neural Network Training – Back Propagation
  • 83. Gradient Descent Algorithm
  • 84. Neural Network Training – Back Propagation
  • 85. Neural Network
  • 86. Neural Network Training – Back Propagation
  • 87. Gradient Descent Algorithm
  • 88. Gradient Descent Algorithm
  • 89. Neural Network — Formal Formulation
  • 90. References for Neural Network
  • 91. Spectrogram
  • 92. Spectrogram
  • 93. Gabor Features (1/2)
  • 94. Gabor Features (2/2)
  • 95. Gabor Features (1/2)
  • 96. Gabor Features (2/2)
  • 97. Gabor Features (1/2)
  • 98. Gabor Features (2/2)
  • 99. Integrating HMM with Neural Networks
  • 100. Gabor Features (2/2)
0/100
Volume
1.0x
00:00/1:10:09