Computer Models of Speech Using Fuzzy Algorithms

It is with great pleasure that I present this third volume of the series "Advanced Applications in Pattern Recognition." It represents the summary of many man- (and woman-) years of effort in the field of speech recognition by tne author's former team at the University of Turin. It combines the best results in fuzzy-set theory and artificial intelligence to point the way to definitive solutions to the speech-recognition problem. It is my hope that it will become a classic work in this field. I take this opportunity to extend my thanks and appreciation to Sy Marchand, Plenum's Senior Editor responsible for overseeing this series, and to Susan Lee and Jo Winton, who had the monumental task of preparing the camera-ready master sheets for publication. Morton Nadler General Editor vii PREFACE Si parva licet componere magnis Virgil, Georgics, 4,176 (37-30 B.C.) The work reported in this book results from years of research oriented toward the goal of making an experimental model capable of understanding spoken sentences of a natural language. This is, of course, a modest attempt compared to the complexity of the functions performed by the human brain. A method is introduced for conce1v1ng modules performing perceptual tasks and for combining them in a speech understanding system.

1. Computer Models for Speech Understanding.- 1.1 Motivations for speech understanding researches.- 1.2 Tasks, difficulties and types of models.- 1.3 A passive model for automatic speech recognition.- 1.4 Active models for speech understanding.- 1.4.1. Elementary psychoacoustic considerations.- 1.4.2. Interpreting the acoustic signal is problem solving.- 1.4.3. Structures for Speech Understanding System models.- 1.4.4. Functional description of the system.- 1.4.5. On the tools for representing and using knowledge.- 1.5 On the use of fuzzy set theory.- 1.6 The structure of the book.- 2. Generation and Recognition of Acoustic Patterns.- 2.1 Speech generation.- 2.2 Techniques for generating acoustic patterns.- 2.2.1. The filter bank.- 2.2.2. The Fast Fourier Transform.- 2.2.3. Identification of the vocal tract parameters.- 2.2.4. Extraction of articulatory parameters.- 2.2.5. On the use of spectral representation of speech.- 2.3 Background on syntactic pattern recognition.- 2.4 Acoustic Cue Extraction for Speech Patterns.- 2.4.1 Silence interval between two sounds in a word.- 2.4.2 Quasi-stationary portions of the acoustic pattern.- 2.4.3 Lines.- 2.5 Classification of speech patterns.- 2.5.1 A brief history of automatic recognition of isolated words.- 2.5.2 The dynamic programming approach.- 2.6 Automatic recognition of continuous speech.- 2.7 References.- 3. On the Use of Syntactic Pattern Recognition and fuzzy Set Theory.- 3.1 Introduction and motivations.- 3.2 The syntactic (structural) approach to the interpretation of speech patterns.- 3.3 The syntax for the recognition of the phonetic feature “vocalic”.- 3.4 Background on fuzzy set theory.- 3.4.1 Definition of fuzzy sets.- 3.4.2 Operations on fuzzy sets.- 3.4.3 Fuzzy restrictions.- 3.4.4 Possibility distributions.- 3.4.5 A simple example.- 3.5 Fuzzy relations and languages.- 3.5.1 Fuzzy relations.- 3.5.2 The extension of principle.- 3.5.3 Fuzzy languages.- 3.6 Use of fuzzy algorithms for feature hypothesization.- 3.6.1 Fuzzy algorithms.- 3.6.2 An example of application.- 3.7 References.- 4. Design Principles for Controlling the Use of Structural Rules for Segmentation.- 4.1 The meaning of the meaning.- 4.2 The control problem in the segmentation process.- 4.3 Computation with linguistic probabilities.- 4.4 Segmentation of continuous speech into pseudo-syllabic nuclei.- 4.4.2 Introduction.- 4.4.2 The segmentation grammar.- 4.4.3 The segmentation algorithm.- 4.4.4 Examples.- 4.5 A parallel processing model for generating phoneme hypotheses.- 4.6 A review of previous work on phoneme recognition.- 4.7 References.- 5. Rules for Characterizing Sonorant Sounds.- 5.1 A fragmant of the structural knowledge source for pseudo-syllables.- 5.1.1 Generalities.- 5.1.2 Generation of hypotheses about sonorant sounds.- 5.2 Extraction of detailed spectral features for sonorant sounds.- 5.2.1 Extraction of a multilinked data structure from a spectrogram.- 5.2.2 Deletion of unsuitable links.- 5.2.3 Assignment of weights to the arcs.- 5.3 Generation of hypotheses about vowels.- 5.3.1 Algorithm SZDET.- 5.3.2 Recognition of the place of articulation of vowels.- 5.3.3 Hypothesis generation and problem solving.- 5.4 Use of formants for the recognition of liquids and nasals.- 5.4.1 Liquid-nasal classification.- 5.4.2 Applications to the classification of liquids.- 5.5 Detailed recognition of nasal sounds.- 5.5.1 Introductory acoustical and perceptual considerations.- 5.5.2 Inference of the recognition rules.- 5.5.2.1 Speech material.- 5.5.2.2 Parameters of the atomic questions.- 5.5.2.3 The recognition rules.- 5.5.3 Experimental results.- 5.5.4 On the extension of the rules to other contexts.- 5.5.5 On the evaluation of binary features.- 5.6 Structure of the procedural knowledge.- 5.7 References.- 6. Rules for Characterizing the Nonsonorant Sounds.- 6.1 Introduction.- 6.2 Recognition of the phonetic features of nonsonorant sounds.- 6.3 Bottom-up generation of phonemic hypotheses of plosive sounds.- 6.3.1 Review of research concerning the plosive consonants.- 6.3.2 Recognition of plosive sounds.- 6.4 Rules for the recognition of plosive sounds.- 6.4.1 Rules for formant loci, formant slopes and burst spectra.- 6.4.2 Rules for spectral characteristics of plosives.- 6.4.3 Rules for formant features.- 6.4.4 Rules for phonemic hypotheses.- 6.4.5 Composition,of evidences.- 6.5 Experimental results.- 6.6 References.- 7. The Lexical Knowledge Source.- 7.1 Word recognition in continuous speech.- 7.2 Dynamic programming for matching word patterns of quasi-continuous feature vectors.- 7.3 Matching speech states.- 7.3.1 Minimum-distance models.- 7.3.2 Stochastic models.- 7.4 Word detection by the hypothesize-and-test paradigm.- 7.5 The lexical component as a problem solver.- 7.6 The structure of the lexical knowledge.- 7.7 Strategies for lexical access.- 7.7.1 Top-down constraints.- 7.7.2 Preconditions based on the first syllable.- 7.7.3 Precondition degradations.- 7.7.4 The lexicon as a content-addressable-memory.- 7.7.5 The syll-type tree.- 7.7.6 Precondition evidences.- 7.7.7 The algorithm for lexical access.- 7.8 Selection of candidates and hypothesis evaluation.- 7.8.1 Evaluation of precondition evidences.- 7.8.2 Candidate selection.- 7.8.3 Other possible methods for hypothesis evaluation.- 7.9 Strategies for the generation of lexical hypotheses.- 7.10 References.- 8. On the Structure and Use of Task-Dependent Knowledge.- 8.1 Introduction.- 8.2 Finite-state language models.- 8.3 Measuring evidences.- 8.4 Search strategies.- 8.4.1 Branch-and-bound algorithms.- 8.4.2 Non-admissible search algorithms.- 8.5 On the use of production systems for problem solving.- 8.6 Scheduling of interpretation processes based on approximate reasoning.- 8.6.1 Background.- 8.6.2 On the use of truth functions and fuzzy logic.- 8.6.3 Priority assignment and approximate reasoning.- 8.7 Outline of a semantically-guided use of task-dependent knowledge.- 8.7.1 System organization.- 8.7.2 The semantic knowledge.- 8.7.3 The syntactic knowledge.- 8.7.4 Pragmatics.- 8.8 Evaluating language complexity.- 8.9 Review of recent work on task-dependent knowledge.- 8.9.1 Representation.- 8.9.2 Control of strategies and scoring philosophies.- 8.10 References.- 9. Automatic Learning of Fuzzy Relations.- 9.1 Introduction.- 9.2 Formal definition of the problem and an example of application.- 9.2.1 Generalities.- 9.2.2 An example of application.- 9.3 A simple preliminary learning case.- 10. Towards a Parallel System.- 10.1 A new model for lexical access.- 10.2 Description of acoustic cues.- 10.3 The knowledge of the descriptor of the global spectral features.- 10.4 Conclusions.