상세정보

Connectionist Speech Recognition: A Hybrid Approach describes the theory and implementation of a method to incorporate neural network approaches into state of the art continuous speech recognition systems based on hidden Markov models (HMMs) to improve their performance. In this framework, neural networks (and in particular, multilayer perceptrons or MLPs) have been restricted to well-defined subtasks of the whole system, i.e. HMM emission probability estimation and feature extraction.
The book describes a successful five-year international collaboration between the authors. The lessons learned form a case study that demonstrates how hybrid systems can be developed to combine neural networks with more traditional statistical approaches. The book illustrates both the advantages and limitations of neural networks in the framework of a statistical systems.
Using standard databases and comparison with some conventional approaches, it is shown that MLP probability estimation can improve recognition performance. Other approaches are discussed, though there is no such unequivocal experimental result for these methods.
Connectionist Speech Recognition is of use to anyone intending to use neural networks for speech recognition or within the framework provided by an existing successful statistical approach. This includes research and development groups working in the field of speech recognition, both with standard and neural network approaches, as well as other pattern recognition and/or neural network researchers. The book is also suitable as a text for advanced courses on neural networks or speech processing.


Connectionist Speech Recognition: A Hybrid Approach describes the theory and implementation of a method to incorporate neural network approaches into state of the art continuous speech recognition systems based on hidden Markov models (HMMs) to improve their performance. In this framework, neural networks (and in particular, multilayer perceptrons or MLPs) have been restricted to well-defined subtasks of the whole system, i.e. HMM emission probability estimation and feature extraction.
The book describes a successful five-year international collaboration between the authors. The lessons learned form a case study that demonstrates how hybrid systems can be developed to combine neural networks with more traditional statistical approaches. The book illustrates both the advantages and limitations of neural networks in the framework of a statistical systems.
Using standard databases and comparison with some conventional approaches, it is shown that MLP probability estimation can improve recognition performance. Other approaches are discussed, though there is no such unequivocal experimental result for these methods.
Connectionist Speech Recognition is of use to anyone intending to use neural networks for speech recognition or within the framework provided by an existing successful statistical approach. This includes research and development groups working in the field of speech recognition, both with standard and neural network approaches, as well as other pattern recognition and/or neural network researchers. The book is also suitable as a text for advanced courses on neural networks or speech processing.

정보제공 :

CONTENTS
List of Figures = xiii
List of Tables = xv
Notation = xix
Foreword = xxiii
Preface = xxv
Acknowledgments = xxvii
Ⅰ BACKGROUND = 1
 1 INTRODUCTION = 3
  1.1 Automatic Speech recognition(ASR) = 4
  1.2 Limitation un Current ASR Systems = 9
  1.3 Book Overview = 10
 2 STATISTICAL PATTERN CLASSIFICATION = 15
  2.1 Introduction = 15
  2.2 A Model or Pattern Classification = 16
  2.3 Statistical Classification = 18
  2.4 Pattern Classification with Realistic Data = 23
  2.5 Summary = 25
 3 HIDDEN MARKOV MODELS = 27
  3.1 Introduction = 27
  3.2 Introduction and Underlying Hypotheses = 31
  3.3 Parametrization and Estimation = 33
   3.3.1 General formulation = 33
   3.3.2 Continuous Input Features = 38
   3.3.3 Discrete Input Features = 39
   3.3.4 Maximum Likelihood Criterion = 40
   3.3.5 Viterbi Criterion = 43
  3.4 Training Problem = 44
   3.4.1 Maximum Likelihood Criterion = 45
   3.4.2 Viterbi Criterion = 47
  3.5 Decoding Problem = 50
   3.5.1 Maximum Likelihood Criterion = 50
   3.5.2 Viterbi Criterion = 51
  3.6 Likelihood and Discrimination = 52
  3.7 Summary = 56
 4 MULTILAYER PERCEPTRONS = 59
  4.1 Introduction = 59
  4.2 Linear Perceptrons = 60
   4.2.1 Linear Discriminant Function = 60
   4.2.2 Least Mean Square Criterion = 61
   4.2.3 Normal Density = 62
  4.3 Multilayer Perceptrons(MLP) = 63
   4.3.1 Some History = 63
   4.3.2 Motivations = 64
   4.3.3 Architecture and Trainning Procedure = 66
   4.3.4 Lagrange Multipliers = 68
   4.3.5 Speeding Up EBP = 71
   4.3.6 On-Line and Off-Line Training = 73
  4.4 Nonlinear Discrimination = 74
   4.4.1 Nonlinear Functions in MLPs = 74
   4.4.2 Phonemic Strings to Words = 76
   4.4.3 Acoustic Vectors to Words = 77
  4.5 MSE and Discriminant Distance = 78
  4.6 Summary = 79
Ⅱ HYBRID HMM／MLP SYSTEMS = 81
 5 SPEECH RECOGNITION USING ANNs = 83
  5.1 Introduction = 83
  5.2 Fallacious Reasons for Using ANNs = 85
  5.3 Valid Reasons for Using ANNs = 87
  5.4 Neural Nets and Time Sequences = 89
   5.4.1 Static Networks with Buffered Input = 90
   5.4.2 Recurrent Networks = 93
   5.4.3 Partial Feed back of Context Units = 94
   5.4.4 Approximating Recurrent Networks by MLPs = 97
   5.4.5 Discussion = 99
  5.5 ANN models of HMMs = 100
   5.5.1 The Viterbi Network = 100
   5.5.2 The Alpha-Net = 101
   5.5.3 Combining ANNs and Dynamic Time Warping = 101
   5.5.4 ANNs for Nonlinear Transformations = 104
   5.5.5 ANNs for Preprocessing = 105
  5.6 Discrimination with Contextual MLPs = 105
  5.7 Summary = 113
 6 STATISTICAL INFERENCE IN MLPs = 115
  6.1 Introduction = 115
  6.2 ANNs and Statistical Inference = 116
   6.2.1 Discrete case = 117
   6.2.2 Continuous Case = 120
  6.3 Recurrent MLP with Output Feedback = 121
  6.4 Practical Implications = 125
   6.4.1 Local Minima = 125
   6.4.2 Network Outputs Sum to One = 125
   6.4.3 Prior Class Probabilities and Likelihoods = 126
   6.4.4 Priors and MLP Output Biases = 127
   6.4.5 Conclusion = 127
  6.5 MLPs with Contextual Inputs = 128
  6.6 Classification of Acoustic Vectors = 130
   6.6.1 Experimental Approach = 130
   6.6.2 MLP Approach, Training and Cross-Validation = 131
   6.6.3 MLP Results = 133
   6.6.4 Assessing Bayesian Properties of MLPs = 133
   6.6.5 Effect of Cross-Validation = 136
   6.6.6 Output Sigmoid Function = 138
   6.6.7 Feature Dependence = 139
  6.7 Radial Basis Functions = 140
   6.7.1 General Approach = 140
   6.7.2 RBFs and Tied Mixtures = 142
   6.7.3 RBFs for MAP Estimation = 143
   6.7.4 Lagrange Multipliers = 145
   6.7.5 Discussion = 147
  6.8 MLPs for Autoregressive Modeling = 147
   6.8.1 Linear Autoregressive Modeling = 148
   6.8.2 Predictive Neural Networks = 149
   6.8.3 Statistical Interpretation = 149
   6.8.4 Another Approach = 150
   6.8.5 Discussion = 151
  6.9 Summary = 152
 7 THE HYBRID HMM／MLP APPROACH = 155
  7.1 Introduction = 155
  7.2 Discriminant Markov Models = 158
   7.2.1 Formulation = 158
   7.2.2 Conditional Transition Probabilities = 159
   7.2.3 Maximum Likelihood Criterion = 162
   7.2.4 Viterbi Criterion = 162
   7.2.5 MLPs for Discriminant HMMs = 163
  7.3 Problem = 165
  7.4 Methods for Recognition at Word Level = 166
   7.4.1 MLP Training Methods = 166
   7.4.2 Posterior Probabilities and Likelihoods = 168
   7.4.3 Word Transition Costs = 168
   7.4.4 Segmentation of Training Data = 170
   7.4.5 Input Features = 170
   7.4.6 Better Speech Units and Phonological Rules = 171
  7.5 Word Recognition Results = 171
  7.6 Segmentation of Training Data = 175
  7.7 Resource Management(RM) task = 177
   7.7.1 Methods = 177
   7.7.2 Results = 179
   7.7.3 Discussion and Extensions = 179
  7.8 Discriminative Training and Priors = 181
  7.9 Summary = 182
 8 EXPERIMENTAL SYSTEMS = 185
  8.1 Introduction = 185
  8.2 Experiments on RM and TIMIT = 187
   8.2.1 Methods = 187
   8.2.2 Recognition Results = 190
   8.2.3 Discussion = 193
  8.3 Integrating the MLP into DECIPHER = 195
   8.3.1 Coming Full Circle : RM Experiments = 196
   8.3.2 Context-independent Models = 197
  8.4 Summary = 199
 9 CONTEXT-DEPENDENT MLPs = 201
  9.1 Introduction = 201
  9.2 CDNN : A Context-Dependent Neural Network = 202
  9.3 Theoretical Issue = 203
  9.4 Implementation Issue = 207
  9.5 Discussion and Results = 209
   9.5.1 The Unrestricted Split Net = 209
   9.5.2 The Topologically Restricted Net = 210
   9.5.3 Preliminary Results and Conclusion = 210
  9.6 Related Prior Work = 211
  9.7 Summary = 212
 10 SYSTEM TRADEOFFS = 215
  10.1 Introduction = 215
  10.2 Discrete HMM = 216
  10.3 Continuous-Density HMM = 218
  10.4 Summary = 221
 11 TRAINING HARDWARE AND SOFTWARE = 223
  11.1 Introduction = 223
  11.2 Motivations = 224
  11.3 A Basic Neurocomputer Design-the RAP = 226
   11.3.1 The ICSI Ring Array Processor = 226
   11.3.2 Current Developments and Conclusions = 228
  11.4 Summary = 229
Ⅲ ADDITIONAL TOPICS = 231
 12 CROSS-VALIDATION IN MLP TRAINING = 233
  12.1 Introduction = 233
  12.2 Random Vector Problem = 235
   12.2.1 Methods = 235
   12.2.2 Results = 236
  12.3 Speech Recognition = 239
   12.3.1 Methods = 239
   12.3.2 Results = 240
  12.4 Summary = 241
 13 HMM／MLP AND PREDICTIVE MODELS = 243
  13.1 Introduction = 243
  13.2 Autoregressive HMMs = 245
  13.3 Full and Conditional Likelihoods = 247
  13.4 Gaussian Additive Noise = 248
   13.4.1 Training = 248
   13.4.2 Recognition = 249
   13.4.3 Discussion = 250
  13.5 Linear or Nonlinear AR Models? = 250
  13.6 ARCH Models = 251
  13.7 Summary = 252
 14 FEATURE EXTRACTION BY MLP = 253
  14.1 Introduction = 253
  14.2 MLP and Auto-Association = 255
  14.3 Explicit and Optimal Solution = 257
  14.4 Linear Hidden Units = 259
  14.5 Nonlinear Hidden Units = 260
  14.6 Experiments = 261
  14.7 Summary = 263
Ⅳ FINALE = 265
 15 FINAL SYSTEM OVERVIEW = 267
  15.1 Introduction = 267
  15.2 System Description = 267
   15.2.1 Network Specifications = 267
   15.2.2 Training = 269
   15.2.3 Recognition = 272
  15.3 New Perspectives = 273
  15.4 Summary = 274
 16 CONCLUSIONS = 275
  16.1 Introduction = 275
  16.2 Hybrid HMM／ANN Systems : Status = 276
  16.3 Future Research Issues = 278
   16.3.1 In Hybrid HMM／ANN Approaches for CSR = 278
   16.3.2 In General = 279
  16.4 Concluding Remark = 280
Bibliography = 281
Index = 307
Acronyms = 311