It is well-known that phonemes have different acoustic realizations depending on the context. Thus, for example, the phoneme /t! is typically realized with a heavily aspirated strong burst at the beginning of a syllable as in the word Tom, but without a burst at the end of a syllable in a word like cat. Variation such as this is often considered to be problematic for speech recogni­ tion: (1) "In most systems for sentence recognition, such modifications must be viewed as a kind of 'noise' that makes it more difficult to hypothesize lexical candidates given an in­ put phonetic transcription. To see that this must be the case, we note that each phonological rule [in a certain example] results in irreversible ambiguity-the phonological rule does not have a unique inverse that could be used to recover the underlying phonemic representation for a lexical item. For example, . . . schwa vowels could be the first vowel in a word like 'about' or the surface realization of almost any English vowel appearing in a sufficiently destressed word. The tongue flap [(] could have come from a /t! or a /d/. " [65, pp. 548-549] This view of allophonic variation is representative of much of the speech recognition literature, especially during the late 1970's. One can find similar statements by Cole and Jakimik [22] and by Jelinek [50].

Inhalt

1. Introduction.- 1.1 Historical Background and Problem Statement.- 1.1.1 Simplifications to the General Speech Understanding Problem.- 1.1.2 Major Components of a CSR System.- 1.2 Allophonic Constraints are Useful.- 1.2.1 "Did you hit it to Tom?".- 1.2.1.1 Phonological Rules: A Source of Noise?.- 1.2.1.2 Phonological Rules: A Source of Constraint.- 1.2.1.3 Redundant Constraints.- 1.2.2 "Plant some more tulips".- 1.2.3 "We make all of our children".- 1.3 Problems with Rewrite-Rules.- 1.4 Trends Toward Larger Constituents.- 1.4.1 Trends in Speech Research.- 1.4.1.1 Advantage 1: Improved Performance due to Sharing.- 1.4.1.2 Advantage 2: Lexicon is Free of Allophones.- 1.4.2 Trends in Linguistic Theory.- 1.4.2.1 The 'Classical' Position.- 1.4.2.2 Metrical Foot Structure.- 1.4.2.2.1 Metrical Foot Structure and Stress.- 1.4.2.2.2 Is Foot Structure Just a Notational Variant?.- 1.5 Parsing and Matching.- 1.5.1 Parsing.- 1.5.2 Matching.- 1.5.2.1 Canonicalization.- 1.5.2.2 The Final Result.- 1.6 Summary.- 1.7 Outline of What's To Come.- 2. Representation of Segments.- 2.1 Stevens' Theory of Invariant Features.- 2.2 Our Position.- 2.3 What's New.- 2.3.1 Use of Parsing Constraints.- 2.3.2 Decompositional View of Segments.- 2.4 Motivations for Representing Phonetic Distinctions.- 2.4.1 Capture Phonetic Constraints on Suprasegmental Constituents.- 2.4.2 Challenge to the Theory of Invariant Features.- 2.4.2.1 Example I: /t, d/.- 2.4.2.2 Example II: Nasals.- 2.4.2.3 Example III: /t, k/.- 2.5 Capturing Generalizations.- 2.5.1 Allophonic Proliferation.- 2.5.2 Consistency in Transcriptions.- 2.5.3 Consistency Among Dictionaries.- 2.6 Summary.- 3. Allophonic Rules.- 3.1 Flapping and Syllable Level Generalizations.- 3.1.1 Modification 1: Optional Consonant.- 3.1.2 Modification 2: Word Boundaries.- 3.1.3 Modification 3: Stress.- 3.2 Non-Linear Formulations of Flapping.- 3.2.1 Kahn's Ambi-Syllabicity.- 3.2.2 Metrical Foot Structure.- 3.2.3 Differences between Syllable and Foot Structure.- 3.2.3.1 Word Boundaries.- 3.2.3.2 Tri-Syllabic Feet.- 3.2.4 The Optional Consonant (Revisited).- 3.2.5 Interaction with Morphology.- 3.2.5.1 #-Prefixes.- 3.2.5.2 #-Suffixes.- 3.2.6 Summary of Argument for Non-Linear Analysis.- 3.3 Implementation Difficulties and the Lexical Expansion Solution.- 3.3.1 Generative in Nature.- 3.3.2 Restrictions on Allophonic Grammars.- 4. An Alternative: Phrase-Structure Rules.- 4.1 PS Trees Bear More Fruit Than You Would Have Thought.- 4.2 The Constituency Hypothesis.- 4.2.1 Foot-Internal Rules.- 4.2.2 Non-Foot-Initial Rules.- 4.2.3 Foot-Initial Rules.- 4.2.4 Re-organization of the Grammar.- 4.3 Advantages of Phrase-Structure Formulation.- 4.3.1 Robustness.- 4.3.2 Efficiency.- 4.3.3 The Chicken or Egg Paradox.- 4.4 Summary.- 5. Parser Implementation.- 5.1 An Introduction to Chart Parsing.- 5.2 Representation Issues.- 5.2.1 The Chart Tends to be Sparse.- 5.2.2 Taking Advantage of the Sparseness.- 5.3 A Parser Based on Matrix Operations.- 5.3.1 Concatenation and Union.- 5.3.2 Optionality.- 5.3.3 Transitive Closure.- 5.4 No Recursion.- 5.5 Order of Evaluation.- 5.6 Feature Manipulation.- 5.7 Additional Lattice Operations.- 5.7.1 Over-generate and Filter.- 5.7.2 Context Primitives.- 5.7.3 Garbage Collection.- 5.8 Debugging Capabilities.- 5.9 Summary.- 6. Phonotactic Constraints.- 6.1 The Affix Position.- 6.2 The Length Restriction.- 6.3 The Sonority Hierarchy.- 6.3.1 Exceptions to the Sonority Hierarchy.- 6.4 Practical Applications of Phonotactic Constraints.- 6.4.1 The Sonority Bound on Syllable Ambiguity.- 6.4.2 Removing Redundancy From the Lexicon.- 6.4.2.1 A Note on Liquids and Glides.- 6.5 Summary.- 7. When Phonotactic Constraints are Not Enough.- 7.1 Basic Principles.- 7.1.1 Maximize Onsets and Stress Resyllabification.- 7.1.2 Morphology.- 7.1.3 Maximize Onsets.- 7.1.4 Stress Resyllabification.- 7.2 Against Stress Resyllabification.- 7.2.1 Alternative 1: No Resyllabification.- 7.2.2 Alternative 2: Limited Stress Resyllabification.- 7.2.3 Alternative 3: Vowel Resyllabification.- 7.2.4 Summary of Alternatives to Stress Resyllabification.- 7.3 Practical Applications of Vowel Resyllabification.- 7.3.1 Alternative Points of View.- 7.3.2 Vowel Class and Allophonic Cues are Often Redundant.- 7.3.3 Reduction of Branching Factor.- 7.3.4 Schwas Enrich the Speech Signal.- 7.3.5 Stronger Form of Vowel Resyllabification.- 7.4 Automatic Syllabification of Lexicons.- 7.5 Summary.- 8. Robustness Issues.- 8.1 Alternatives in the Input Lattice.- 8.2 Problems for Parsing.- 8.2.1 Allophonic Constraints on an Impoverished Lattice.- 8.2.2 Phonotactic Constraints Aren't Helping Either.- 8.3 Relaxing Phonological Distinctions.- 8.4 Conservation of Distinctive Features.- 8.4.1 Deletion in Homorganic Nasal Clusters.- 8.4.2 Deletion in Fricative Clusters.- 8.4.3 Tunafish Sandwich.- 8.5 Probabilistic Methods.- 8.5.1 Problems with Relaxing Distinctions.- 8.5.2 The Similarity of Probabilistic and Categorical Approaches.- 8.6 Distinctive Features.- 8.6.1 Modeling Confusions at the Segmental Level.- 8.6.2 A Comparison of Seneff's Performance with BBN's Front End.- 8.6.3 Practical Application of Confusion Matrices.- 8.6.4 Modelling Confusions at the Distinctive Feature Level.- 8.6.5 Feature Integration.- 8.6.6 Practical Applications.- 8.7 Summary.- 9. Conclusion.- 9.1 Review of the Standard Position.- 9.2 Review of Nakatani's Position.- 9.3 Review of the Constituency Hypothesis.- 9.4 Review of Phonotactic Constraints.- 9.5 Comparison with Syntactic Notions of Constituency.- 9.6 Contributions.- References.- Appendix I. The Organization of the Lexicon.- I.1. Linear Representation and Linear Search.- I.2. Non-Recursive Discrimination Networks.- I.3. Recursive Discrimination Networks.- I.4. Hash Tables Based on Equivalence Class Abstractions.- I.5. Shipman and Zue.- I.6. Morse Code.- I.7. Selecting the Appropriate Gross Classification.- I.8. Summary.- Appendix II. Don't Depend Upon Syntax and Semantics.- II.1. Higher Level vs. Lower Level Constraints.- II.2. Too Much Dependence in the Past.- II.3. How Much Can Higher Constraints Help?.- II.4. Detraction from the Important Low-Level Issues.- II.5. New Directions: Recognition without Understanding.- II.6. Lower-Level Constraints Bear More Fruit.- II.7. Summary.- Appendix III. Lexical Phonology.- III.1. Difference Between + and #.- III.2. Pipeline Design.- III.3. Distinctions Between Lexical and Postlexical Rules.- III.4. Which Rules are Lexical and Which are Postlexical?.- III.5. The Implementation of Lexical and Postlexical Rules.- Appendix IV. A Sample Grammar.- Appendix V. Sample Lexicon.- Appendix VI. Sa…