Speech Recognition Theory and C++ Implementation

by Becchetti, Claudio; Ricotti, Lucio Prina

• ISBN13: 9780471977308
• ISBN10: 0471977306
• Edition: 1st
• Format: Hardcover
• Copyright: 1999-05-04
• Publisher: WILEY

Summary

Automatic Speech Recognition (ASR) is the enabling technology for hands-free dictation and voice-triggered computer menus. It is becoming increasingly prevalent in environments such as private telephone exchanges and real-time information services. Speech Recognition introduces the principles of ASR systems, including the theory and implementation issues behind multi-speaker continuous speech recognition. Focusing on the algorithms employed in commercial and laboratory systems, the treatment enables the reader to devise practical solutions for ASR system problems. It addresses in detail C++ programming techniques used to develop ASR applications, thus offering skills that will prove useful in any large C++ based software project. Possible extensions of the well-established ASR technology are highlighted, based on "Hidden Markov Models" applied to fields such as modelling and prediction of econometric series. Features include: Accompanying CD-ROM containing all C++ source code of a complete laboratory multi-speaker continuous-speech ASR system (e.g. Initialisation, Training, Recognition, Evaluation, etc.) Detailed theoretical, mathematical and technical explanations of ASR A practical account of the functioning of ASR A crucial source of information for researchers, developers and project managers involved with ASR systems, Speech Recognition is also structured for use by students of digital signal processing, speech recognition and C++ programming techniques.

Author Biography

Claudio Becchetti graduated with honors in Electronic Engineering in 1994 at the University of Rome, where he achieved the Ph.D. in Telecommunications in 1999. From 2002 to 2009, he was adjoint professor at the University "La Sapienza", faculty of Telecommunication Engineering where he held first a course on Industrial design and then a course on Signal Theory. Claudio has 7 years teaching experience working with students studying ECG. This device is well suited as a practical example for signal theory, digital signal processing, electronics and software engineering.

Table of Contents

Preface	xv
Acknowledgments	xviii
Statistical Speech Recognition	1	(66)
Part I: Theory
Introduction	2	(4)
The theoretical part	3	(2)
The implementation part	5	(1)
ASR: A Simple Explanation abc	6	(2)
Statistical Speech Recognition abc	8	(10)
The deterministic approach	9	(1)
The stochastic framework	10	(1)
Stochastic model simplifications	11	(3)
Hidden Markov Models: an introduction	14	(3)
The search algorithm	17	(1)
ASR Structure and Chapter Organization ***	18	(2)
Problems, Applications and Products in ASR	20	(2)
The ASR as an Information Theory Problem ***	22	(6)
The layered model	22	(2)
The L-language	24	(1)
The information theory approach	24	(4)
Part II: Implementation
Introduction	28	(3)
The Crisis of Software Development *** ***	31	(2)
Object oriented versus procedural programming	32	(1)
A Safe C++ Programming Style ***	33	(10)
What must not be used in C++ ***	35	(3)
Syntax and programming style ***	38	(1)
Conventions in writing code ***	39	(2)
Basic design rules: the modules ***	41	(2)
The Program Services ***	43	(2)
The Memory Service ***	45	(10)
Dynamic memory management: a C++ without pointers ***	46	(4)
Object list resizing: the C++ ``realloc'' ***	50	(1)
The root class: a user guide	51	(3)
Implementation details of the container classes ***	54	(1)
Diagnostics ***	55	(6)
Defensive programming ***	55	(3)
The diagnostic system (User guide)	58	(2)
Implementation details of the diagnostic system ***	60	(1)
Topics Related to Large Projects *** ***	61	(6)
Portability and defined options (file compatib.h defopt.h Boolean.h)	61	(2)
Test routines and version history	63	(1)
Time efficiency, development speed and code readability	64	(3)
Speech Database	67	(54)
Part I: Theory
Introduction	68	(1)
Statistical Notes	69	(1)
Speech Variability ***	70	(4)
Design of Speech Databases ***	74	(8)
Some practical aspects of building a corpus	78	(2)
Transcriptions of speech	80	(2)
Speech Database Sources	82	(6)
TIMIT and ATIS databases ***	83	(5)
Part II: Implementation
Introduction	88	(2)
The Configuration Manager ***	90	(5)
The configuration file ***	91	(1)
The option configuration file *.opt ***	92	(1)
The configuration class ***	93	(2)
Introducing new parameters ***	95	(1)
The String Class ***	95	(5)
The String class usage ***	96	(1)
String class techniques ***	97	(3)
The Structure of an ASR-Oriented Database	100	(4)
The options of the database manager ***	103	(1)
The Database Manager Class Hierarchy	104	(9)
The sound file hierarchy (file soundfil.h)	105	(2)
The label file hierarchy	107	(3)
Soundlabelled File and Dbase Voc classes (file soundlab.*)	110	(3)
Safe Polymorphism *** ***	113	(5)
Implementation	115	(3)
Adding new database standards ***	118	(1)
Hints on Class Design *** ***	118	(3)
Speech Signal Analysis	121	(46)
Part I: Theory
Introduction	122	(1)
Analog to Digiral Voice Conversion ***	122	(3)
Physical features of speech signals ***	124	(1)
Feature Extraction ***	125	(9)
Signal preprocessing	125	(1)
Windowing	126	(3)
Spectral analysis	129	(1)
Filter bank processing	130	(2)
Log energy computation	132	(1)
Mel frequency cepstrum computation	132	(2)
Delta coefficients and energy	134	(1)
Cepstrum Analysis	134	(2)
Robustness in Speech Recognition ***	136	(5)
Additive noise and linear distortion model	137	(1)
Cepstral compensation	138	(2)
Channel equalization and speech enhancement	140	(1)
Model adaptation techniques	141	(1)
Distortion Measures ***	141	(3)
Part II: Implementation
Introduction	144	(2)
Feature Extraction: a DSP Approach ***	146	(3)
Configuration of the feature manager	148	(1)
Hierarchy Design	149	(5)
Adding a new block ***	153	(1)
Safe Interfacing C-Style Routines *** ***	154	(4)
The Mathematical Classes ***	158	(9)
Overview of mathematical classes ***	160	(1)
Mathematical vectors in C++ ***	161	(2)
Advanced techniques for vector implementation ***	163	(4)
HMMs and Initialization	167	(44)
Part I: Theory
Introduction	168	(1)
Markov Chains ***	168	(2)
Hidden Markov Models ***	170	(5)
Alternative definitions for HMMs	171	(1)
Processes suitable for HMM modeling	171	(1)
Example: two state-HMM	172	(3)
Application of the HMM to Speech Modeling	175	(9)
Probability inference in HMMs: forward and backward procedure	179	(1)
Observation densities	180	(2)
HMM initialization	182	(2)
Automatic Segmentation	184	(1)
Clustering of Feature Vectors	185	(3)
HMMs versus Probabilistic, Bayesian and Neural Networks ***	188	(4)
Part II: Implementation
Introduction	192	(1)
HMM Phoneme Models	193	(1)
HMM Estimation Procedure ***	194	(4)
Implementation Details	198	(4)
HMM model implementation	198	(2)
Algorithm implementation	200	(2)
Sectionate	202	(1)
Module Options	202	(2)
Array Class ***	204	(3)
Array class regerence guide ***	205	(2)
Diagonal Arrays ***	207	(4)
Diagonal Array user guide	207	(1)
Diagonal array: implementation details ***	208	(3)
HMM Training	211	(32)
Part II: Theory
Introduction	212	(1)
MLE and MAP Training	213	(1)
Utterance Probabilities and HMM Parameters	214	(1)
The Maximization Procedure: MLE Case	215	(6)
Parameter estimation formulas for MLE ***	217	(2)
Summary of MLE procedure	219	(2)
Maximum a Posteriori (MAP) Estimation	221	(7)
Explicit formulas for MAP	223	(1)
Estimate of the hyper-parameters	223	(1)
Use of MAP procedure	224	(4)
MAP Applications	228	(1)
Part II: Implementation
Introduction	229	(1)
Single Model Training	230	(4)
Single model re-estimation algorithm	231	(3)
Model Simultaneous Training	234	(4)
Algorithm description	234	(2)
Algorithm optimizations	236	(2)
Logarithmic Math	238	(1)
Class Organization	238	(1)
Labels: Hints for Futher Developments	239	(1)
Option Description	240	(3)
Language Models	243	(60)
Part I: Theory
Introduction	244	(1)
The Language Models	245	(1)
Languages as Information Sources	246	(3)
The M-gram Model	249	(2)
Perplexity	251	(1)
Rare Events Estimation: Smoothing	252	(7)
Good-Turing smoothing	254	(2)
Linear interpolation model	256	(1)
Non-linear interpolation	257	(2)
Smoothing Methods: Examples	259	(7)
Perplexity: an example	265	(1)
Word Automatic Clustering ***	266	(7)
Estimation accuracy of class-based and word-based probabilities	267	(3)
Clustering algorithms	270	(1)
Equivalent word merging algorithm	270	(3)
Adaptive Language Models ***	273	(1)
Appendix	274	(3)
Part II: Implementation
Introduction	277	(2)
Language Training Model Options	279	(2)
Hierarchy Design	281	(3)
Database Interface and Counts	284	(11)
Segmented database (TIMIT)	285	(3)
Non-segmented databases	288	(4)
Sorted list and binary search ***	292	(3)
Clustering	295	(5)
Merging algorithm	297	(2)
Perplexity computation for class-based language models	299	(1)
Transition Probability Computation	300	(3)
Recognition	303	(38)
Part I: Theory
Introduction	304	(1)
Static Structure of the Recognizer ***	305	(4)
Static structure of a phonetic recognizer	306	(1)
Static structure of a word recognizer	307	(2)
The Search Algorithms ***	309	(4)
The Viterbi algorithm	310	(1)
Algorithm description	311	(2)
Dynamic Structure Implementation ***	313	(2)
Example of the Viterbi Algorithm	315	(5)
Use of the logarithms	320	(1)
Variants of the Static Structures ***	320	(6)
Part II: Implementation
Introduction	326	(1)
Static Structure	327	(6)
Graph implementation *** ***	328	(1)
Layered graph	329	(4)
Dynamic Structure	333	(5)
Memory management	333	(2)
Representing trees	335	(1)
Hypothesis management	336	(2)
Recognition Options	338	(3)
Evaluation and Parameter Setting	341	(22)
Part I: Theory
Introduction	342	(3)
Evaluation conditions	343	(2)
Speech Data Retrieval	345	(1)
Feature Parameters	345	(1)
Initialization	346	(2)
Training	348	(4)
Grammar	352	(1)
Phonetic grammar	352	(1)
Word language model	353	(1)
Recongnition	353	(6)
Phonetic recognition	354	(2)
Word recognition	356	(3)
Part II: Implementation
Evaluation	359	(1)
Options of the evaluator	359	(1)
RES Specifications	360	(3)
Econometric Appendix: The Behavior of Financial Time Series
Introduction	363	(1)
The state of the art: the homogeneous rational expectation hypothesis approach	364	(4)
Descriptive evidence from ``blue chips'' in Italy. USA. UK and Germany	368	(2)
Autoregressive conditional heteroskedasticity models for stock returns	370	(2)
Research frontiers on statistical properties of stock returns: the HMM approach	372	(1)
Conclusions	373	(1)
Empirical results	374	(1)
Diagnostics	374	(13)
References	387	(12)
Index	399