Mathematical models for speech technology / Stephen E. Levinson.

Format:

Book

Author/Creator:

Levinson, Stephen C.

Contributor:

Classes of 1883 and 1884 Fund.

Language:

English

Subjects (All):

Speech processing systems.

Computational linguistics.

Applied linguistics--Mathematics.

Applied linguistics.

Stochastic processes.

Knowledge, Theory of.

Mathematics.

Physical Description:

xx, 261 pages : illustrations ; 26 cm

Place of Publication:

Chichester, West Sussex, England ; Hoboken, NJ, USA : John Wiley, [2005]

Summary:

Humans use language to convey meaningful messages to each other. Linguistic competence consists in the ability to express meaning reliably, not simply to obtain faithful lexical transcriptions. This invaluable reference tool is the product of many years' experience and research on language and speech technology. It presents the motivations for, intuitions behind and basic mathematical models of natural spoken language communication. From a preliminary discussion on the physics of speech production and the taxonomy of linguistic structure, there is a natural progression taking in issues of grammatical inference, automatic speech recognition and constructive theories of language. The author counterbalances theoretical explanations and illustrations with questions of a more philosophical nature designed to highlight the seemingly limitless future potential of speech technology.

Contents:

1.1 Milestones in the history of speech technology 1

1.2 Prospects for the future 3

1.3 Technical synopsis 4

2.1 The physics of speech production 9

2.1.1 The human vocal apparatus 9

2.1.2 Boundary conditions 14

2.1.3 Non-stationarity 16

2.1.4 Fluid dynamical effects 16

2.2 The source-filter model 17

2.3 Information-bearing features of the speech signal 17

2.3.1 Fourier methods 19

2.3.2 Linear prediction and the Webster equation 21

2.4 Time-frequency representations 23

2.5 Classification of acoustic patterns in speech 27

2.5.1 Statistical decision theory 28

2.5.2 Estimation of class-conditional probability density functions 30

2.5.3 Information-preserving transformations 39

2.5.4 Unsupervised density estimation - quantization 42

2.5.5 A note on connectionism 43

2.6 Temporal invariance and stationarity 44

2.6.1 A variational problem 45

2.6.2 A solution by dynamic programming 47

2.7 Taxonomy of linguistic structure 51

2.7.1 Acoustic phonetics, phonology, and phonotactics 52

2.7.2 Morphology and lexical structure 55

2.7.3 Prosody, syntax, and semantics 55

2.7.4 Pragmatics and dialog 56

3 Mathematical models of linguistic structure 57

3.1 Probabilistic functions of a discrete Markov process 57

3.1.1 The discrete observation hidden Markov model 57

3.1.2 The continuous observation case 80

3.1.3 The autoregressive observation case 87

3.1.4 The semi-Markov process and correlated observations 88

3.1.5 The non-stationary observation case 99

3.1.6 Parameter estimation via the EM algorithm 107

3.1.7 The Cave-Neuwirth and Poritz results 107

3.2 Formal grammars and abstract automata 109

3.2.1 The Chomsky hierarchy 110

3.2.2 Stochastic grammars 113

3.2.3 Equivalence of regular stochastic grammars and discrete HMMs 114

3.2.4 Recognition of well-formed strings 115

3.2.5 Representation of phonology and syntax 116

4 Syntactic analysis 119

4.1 Deterministic parsing algorithms 119

4.1.1 The Dijkstra algorithm for regular languages 119

4.1.2 The Cocke-Kasami-Younger algorithm for context-free languages 121

4.2 Probabilistic parsing algorithms 122

4.2.1 Using the Baum algorithm to parse regular languages 122

4.2.2 Dynamic programming methods 123

4.2.3 Probabilistic Cocke-Kasami-Younger methods 130

4.2.4 Asynchronous methods 130

4.3 Parsing natural language 131

4.3.1 The right-linear case 132

4.3.2 The Markovian case 133

4.3.3 The context-free case 133

5 Grammatical Inference 137

5.1 Exact inference and Gold's theorem 137

5.2 Baum's algorithm for regular grammars 137

5.3 Event counting in parse trees 139

5.4 Baker's algorithm for context-free grammars 140

6 Information-theoretic analysis of speech communication 143

6.1 The Miller et al. experiments 143

6.2 Entropy of an information source 143

6.2.1 Entropy of deterministic formal languages 144

6.2.2 Entropy of languages generated by stochastic grammars 150

6.2.3 Epsilon representations of deterministic languages 153

6.3 Recognition error rates and entropy 153

6.3.1 Analytic results derived from the Fano bound 154

6.3.2 Experimental results 156

7 Automatic speech recognition and constructive theories of language 157

7.1 Integrated architectures 157

7.2 Modular architectures 161

7.2.1 Acoustic-phonetic transcription 161

7.2.2 Lexical access 162

7.2.3 Syntax analysis 165

7.3 Parameter estimation from fluent speech 166

7.3.1 Use of the Baum algorithm 166

7.3.2 The role of text analysis 167

7.4 System performance 168

7.5 Other speech technologies 169

7.5.1 Articulatory speech synthesis 169

7.5.2 Very low-bandwidth speech coding 170

7.5.3 Automatic language identification 170

7.5.4 Automatic language translation 171

8 Automatic speech understanding and semantics 173

8.1 Transcription and comprehension 173

8.2 Limited domain semantics 174

8.2.1 A semantic interpreter 175

8.2.2 Error recovery 182

8.3 The semantics of natural language 189

8.3.1 Shallow semantics and mutual information 189

8.3.2 Graphical methods 190

8.3.3 Formal logical models of semantics 190

8.3.4 Relationship between syntax and semantics 194

8.4 System architectures 195

8.5 Human and machine performance 197

9 Theories of mind and language 199

9.1 The challenge of automatic natural language understanding 199

9.2 Metaphors for mind 199

9.2.1 Wiener's cybernetics and the diachronic history 201

9.2.2 The crisis in the foundations of mathematics 205

9.2.3 Turing's universal machine 210

9.2.4 The Church-Turing hypothesis 212

9.3 The artificial intelligence program 213

9.3.1 Functional equivalence and the strong theory of AI 213

9.3.2 The broken promise 214

9.3.3 Schorske's causes of cultural decline 214

9.3.4 The ahistorical blind alley 215

9.3.5 Observation, introspection and divine inspiration 215

9.3.6 Resurrecting the program by unifying the synchronic and diachronic 216

10 A Speculation on the prospects for a science of mind 219

10.1 The parable of the thermos bottle: measurements and symbols 219

10.2 The four questions of science 220

10.2.1 Reductionism and emergence 220

10.2.2 From early intuition to quantitative reasoning 221

10.2.3 Objections to mathematical realism 223

10.2.4 The objection from the diversity of the sciences 224

10.2.5 The objection from Cartesian duality 225

10.2.6 The objection from either free will or determinism 225

10.2.7 The postmodern objection 226

10.2.8 Beginning the new science 227

10.3 A constructive theory of mind 228

10.3.1 Reinterpreting the strong theory of AI 228

10.3.2 Generalizing the Turing test 228

10.4 The problem of consciousness 229

10.5 The role of sensorimotor function, associative memory and reinforcement learning in automatic acquisition of spoken language by an autonomous robot 230

10.5.1 Embodied mind from integrated sensorimotor function 231

10.5.2 Associative memory as the basis for thought 231

10.5.3 Reinforcement learning via interaction with physical reality 232

10.5.4 Semantics as sensorimotor memory 234

10.5.5 The primacy of semantics in linguistic structure 234

10.5.6 Thought as linguistic manipulation of mental representations of reality 235

10.5.7 Illy the autonomous robot 235

10.5.8 Software 237

10.5.9 Associative memory architecture 238

10.5.10 Performance 238

10.5.11 Obstacles to the program 239

10.6 Final thoughts: predicting the course of discovery 241.

Notes:

Includes bibliographical references (pages [243]-255) and index.

Local Notes:

Acquired for the Penn Libraries with assistance from the Classes of 1883 and 1884 Fund.

ISBN:

0470844078

OCLC:

57007719