Digital Speech: Coding for Low Bit Rate Communication Systems / Edition 2

by A. M. Kondoz

Overview

Building on the success of the first edition Digital Speech offers extensive new, updated and revised material based upon the latest research. This Second Edition continues to provide the fundamental technical background required for low bit rate speech coding and the hottest developments in digital speech coding techniques that are applicable to evolving communication systems.

• Features new chapters on Pitch Estimation and Voice-Unvoiced Classification of Speech, Harmonic Speech Coding and Multimode Speech Coding
• Presents a comprehensively revised chapter entitled Analysis by Synthesis LPC Coding including specific examples of popular speech coders such as CELP (Code-Excited Linear Predictive) Coding
• Contains an updated chapter on Efficient LPC Quantization Methods including MSVQ and anti-aliasing filtering
• Discusses Voice Activity Detection (VAD) methods
• Offers expanded coverage of speech enhancement techniques such as echo cancellation and noise suppression

Written by a well-known, highly respected academic, this authoritative volume will be invaluable to practising engineers, network designers, computer scientists and advanced students in communications, electrical and electronic engineering.

Product Details

ISBN-13:	9780470870082
Publisher:	Wiley
Publication date:	10/29/2004
Edition description:	REV
Pages:	464
Product dimensions:	6.61(w) x 9.61(h) x 1.06(d)

About the Author

Professor Kondoz joined the university of Surrey as a PhD. student in October 1984. From 1986 to 1988 he was employed as a research fellow in the communications group. After completing his PhD, in 1988 he was appointed as a lecturer. In 1995 he became a Reader and in 1997 Professor and Deputy Director in the Centre for Communication Systems Research (CCSR).
He has been involved in teaching of digital signal processing, telecommunications theory and source coding in both undergradute and postgraduate levels. In research he has been heading Multimedia Communication Research Group since 1990. To date, Professor Kondoz has supervised 20 successful PhD students in Speech, Video and Channel coding, Source data packetisation, Error resilient speech and video transmission and Mobile multimedia communications. His current research interests are, Low bit rate speech, image and video coding error resilient video transmission, mobile multimedia communications, robust wireless ATM, real-time terminal design and implementation for mobile communications.
Outside the University, Professor Kondoz has been a member of both the IEE and IEEE. He is a CEng and served on E5. He is on EPSRC College for signal processing and communications.

Read an Excerpt

Digital Speech

---

By A.M. Kondoz

John Wiley & Sons, Inc.

Copyright © 2004 John Wiley & Sons, Ltd
All right reserved.
ISBN: 0-470-87009-5

---

Chapter One

Introduction

Although data links are increasing in bandwidth and are becoming faster, speech communication is still the most dominant and common service in telecommunication networks. The fact that commercial and private usage of telephony in its various forms (especially wireless) continues to grow even a century after its first inception is obvious proof of its popularity as a form of communication. This popularity is expected to remain steady for the foreseeable future. The traditional plain analogue system has served telephony systems remarkably well considering its technological simplicity. However, modern information technology requirements have introduced the need for a more robust and flexible alternative to the analogue systems. Although the encoding of speech other than straight conversion to an analogue signal has been studied and employed for decades, it is only in the last 20 to 30 years that it has really taken on significant prominence. This is a direct result of many factors, including the introduction of many new application areas.

The attractions of digitally-encoded speech are obvious. As speech is condensed to a binary sequence, all of the advantages offered by digital systems are available for exploitation. These include the ease of regeneration and signalling, flexibility, security, and integration into the evolving new wireless systems. Although digitally-encoded speech possesses many advantages over its analogue counterpart, it nevertheless requires extra bandwidth for transmission if it is directly applied (without compression). The 64 kb/s Log-PCM and 32 kb/s ADPCM systems which have served the many early generations of digital systems well over the years have therefore been found to be inadequate in terms of spectrum efficiency when applied to the new, bandwidth limited, communication systems, e.g. satellite communications, digital mobile radio systems, and private networks. In these and other systems, the bandwidth and power available is severely restricted, hence signal compression is vital. For digitized speech, the signal compression is achieved via elaborate digital signal processing techniques that are facilitated by the rapid improvement in digital hardware which has enabled the use of sophisticated digital signal processing techniques that were not feasible before. In response to the requirement for speech compression, feverish research activity has been pursued in all of the main research centres and, as a result, many different strategies have been developed for suitably compressing speech for bandwidth-restricted applications. During the last two decades, these efforts have begun to bear fruit. The use of low bit-rate speech coders has been standardized in many international, continental and national communication systems. In addition, there are a number of private network operators who use low bit-rate speech coders for specific applications.

The speech coding technology has gone through a number of phases starting with the development and deployment of PCM and ADPCM systems. This was followed by the development of good quality medium to low bit-rate coders covering the range from 16 kb/s to 8 kb/s. At the same time, very low bit-rate coders operating at around 2.4 kb/s produced better quality synthetic speech at the expense of higher complexity. The latest trend in speech coding is targeting the range from about 6 kb/s down to 2 kb/s by using speech-specific coders, which rely heavily on the extraction of speech-specific information from the input source. However, as the main applications of the low to very low bit-rate coders are in the area of mobile communication systems, where there may be significant levels of background noise, the accurate determination of the speech parameters becomes more difficult. Therefore the use of active noise suppression as a preprocessor to low bit-rate speech coding is becoming popular.

In addition to the required low bit-rate for spectral efficiency, the cost and power requirements of speech encoder/decoder hardware are very important. In wireless personal communication systems, where hand-held telephones are used, the battery consumption, cost and size of the portable equipment have to be reasonable in order to make the product widely acceptable.

In this book an attempt is made to cover many important aspects of low bit-rate speech coding. In Chapter 2, the background to speech coding, including the existing standards, is discussed. In Chapter 3, after briefly reviewing the sampling theorem, scalar and vector quantization schemes are discussed and formulated. In addition, various quantization types which are used in the remainder of this book are described.

In Chapter 4, speech analysis and modelling tools are described. After discussing the effects of windowing on the short-time Fourier transform of speech, extensive treatment of short-term linear prediction of speech is given. This is then followed by long-term prediction of speech. Finally, pitch detection methods, which are very important in speech vocoders, are discussed.

It is very important that the quantization of the linear prediction coefficients (LPC) of low bit-rate speech coders is performed efficiently both in terms of bit rate and sensitivity to channel errors. Hence, in Chapter 5, efficient quantization schemes of LPC parameters in the form of Line Spectral Frequencies are formulated, tested and compared.

In Chapter 6, more detailed modelling/classification of speech is studied. Various pitch estimation and voiced - unvoiced classification techniques are discussed.

In Chapter 7, after a general discussion of analysis by synthesis LPC coding schemes, code-excited linear prediction (CELP) is discussed in detail.

In Chapter 8, a brief review harmonic coding techniques is given.

In Chapter 9, a novel hybrid coding method, the integration of CELP and harmonic coding to form a multi-modal coder, is described.

Chapters 10 and 11 cover the topics of voice activity detection and speech enhancements methods, respectively.

(Continues...)

---

Excerpted from Digital Speech by A.M. Kondoz Copyright © 2004 by John Wiley & Sons, Ltd.
Excerpted by permission.
All rights reserved. No part of this excerpt may be reproduced or reprinted without permission in writing from the publisher.
Excerpts are provided by Dial-A-Book Inc. solely for the personal use of visitors to this web site.

---

Table of Contents

Preface xiii

Acknowledgements xv

1 Introduction 1

2 Coding Strategies and Standards 5

2.1 Introduction 5

2.2 Speech Coding Techniques 6

2.2.1 Parametric Coders 7

2.2.2 Waveform-approximating Coders 8

2.2.3 Hybrid Coding of Speech 8

2.3 Algorithm Objectives and Requirements 9

2.3.1 Quality and Capacity 9

2.3.2 Coding Delay 10

2.3.3 Channel and Background Noise Robustness 10

2.3.4 Complexity and Cost 11

2.3.5 Tandem Connection and Transcoding 11

2.3.6 Voiceband Data Handling 11

2.4 Standard Speech Coders 12

2.4.1 ITU-T Speech Coding Standard 12

2.4.2 European Digital Cellular Telephony Standards 13

2.4.3 North American Digital Cellular Telephony Standards 14

2.4.4 Secure Communication Telephony 14

2.4.5 Satellite Telephony 15

2.4.6 Selection of a Speech Coder 15

2.5 Summary 18

Bibliography 18

3 Sampling and Quantization 23

3.1 Introduction 23

3.2 Sampling 23

3.3 Scalar Quantization 26

3.3.1 Quantization Error 27

3.3.2 Uniform Quantizer 28

3.3.3 Optimum Quantizer 29

3.3.4 Logarithmic Quantizer 32

3.3.5 Adaptive Quantizer 33

3.3.6 Differential Quantizer 36

3.4 Vector Quantization 39

3.4.1 Distortion Measures 42

3.4.2 Codebook Design 43

3.4.3 Codebook Types 44

3.4.4 Training, Testing and Codebook Robustness 52

3.5 Summary 54

Bibliography 54

4 Speech Signal Analysis and Modelling 57

4.1 Introduction 57

4.2 Short-Time Spectral Analysis 57

4.2.1 Role of Windows 58

4.3 Linear Predictive Modelling of Speech Signals 65

4.3.1 Source Filter Model of Speech Production 65

4.3.2 Solutions to LPC Analysis 67

4.3.3 Practical Implementation of the LPC Analysis 74

4.4 Pitch Prediction 77

4.4.1 Periodicity in Speech Signals 77

4.4.2 Pitch Predictor (Filter) Formulation 78

4.5 Summary 84

Bibliography 84

5 Efficient LPC Quantization Methods 87

5.1 Introduction 87

5.2 Alternative Representation of LPC 87

5.3 LPC to LSF Transformation 90

5.3.1 Complex Root Method 95

5.3.2 Real Root Method 95

5.3.3 Ratio Filter Method 98

5.3.4 Chebyshev Series Method 100

5.3.5 Adaptive Sequential LMS Method 100

5.4 LSF to LPC Transformation 101

5.4.1 Direct Expansion Method 101

5.4.2 LPC Synthesis Filter Method 102

5.5 Properties of LSFs 103

5.6 LSF Quantization 105

5.6.1 Distortion Measures 106

5.6.2 Spectral Distortion 106

5.6.3 Average Spectral Distortion and Outliers 107

5.6.4 MSE Weighting Techniques 107

5.7 Codebook Structures 110

5.7.1 Split Vector Quantization 111

5.7.2 Multi-Stage Vector Quantization 113

5.7.3 Search strategies for MSVQ 114

5.7.4 MSVQ Codebook Training 116

5.8 MSVQ Performance Analysis 117

5.8.1 Codebook Structures 117

5.8.2 Search Techniques 117

5.8.3 Perceptual Weighting Techniques 119

5.9 Inter-frame Correlation 121

5.9.1 LSF Prediction 122

5.9.2 Prediction Order 124

5.9.3 Prediction Factor Estimation 125

5.9.4 Performance Evaluation of MA Prediction 126

5.9.5 Joint Quantization of LSFs 128

5.9.6 Use of MA Prediction in Joint Quantization 129

5.10 Improved LSF Estimation Through Anti-Aliasing Filtering 130

5.10.1 LSF Extraction 131

5.10.2 Advantages of Low-pass Filtering in Moving Average Prediction 135

5.11 Summary 146

Bibliography 146

6 Pitch Estimation and Voiced–Unvoiced Classification of Speech 149

6.1 Introduction 149

6.2 Pitch Estimation Methods 150

6.2.1 Time-Domain PDAs 151

6.2.2 Frequency-Domain PDAs 155

6.2.3 Time- and Frequency-Domain PDAs 158

6.2.4 Pre- and Post-processing Techniques 166

6.3 Voiced–Unvoiced Classification 178

6.3.1 Hard-Decision Voicing 178

6.3.2 Soft-Decision Voicing 189

6.4 Summary 196

Bibliography 197

7 Analysis by Synthesis LPC Coding 199

7.1 Introduction 199

7.2 Generalized AbS Coding 200

7.2.1 Time-Varying Filters 202

7.2.2 Perceptually-based Minimization Procedure 203

7.2.3 Excitation Signal 206

7.2.4 Determination of Optimum Excitation Sequence 208

7.2.5 Characteristics of AbS-LPC Schemes 212

7.3 Code-Excited Linear Predictive Coding 219

7.3.1 LPC Prediction 221

7.3.2 Pitch Prediction 222

7.3.3 Multi-Pulse Excitation 230

7.3.4 Codebook Excitation 238

7.3.5 Joint LTP and Codebook Excitation Computation 252

7.3.6 CELP with Post-Filtering 255

7.4 Summary 258

Bibliography 258

8 Harmonic Speech Coding 261

8.1 Introduction 261

8.2 Sinusoidal Analysis and Synthesis 262

8.3 Parameter Estimation 263

8.3.1 Voicing Determination 264

8.3.2 Harmonic Amplitude Estimation 266

8.4 Common Harmonic Coders 268

8.4.1 Sinusoidal Transform Coding 268

8.4.2 Improved Multi-Band Excitation, INMARSAT-M Version 270

8.4.3 Split-Band Linear Predictive Coding 271

8.5 Summary 275

Bibliography 275

9 Multimode Speech Coding 277

9.1 Introduction 277

9.2 Design Challenges of a Hybrid Coder 280

9.2.1 Reliable Speech Classification 281

9.2.2 Phase Synchronization 281

9.3 Summary of Hybrid Coders 281

9.3.1 Prototype Waveform Interpolation Coder 282

9.3.2 Combined Harmonic and Waveform Coding at Low Bit-Rates 282

9.3.3 A 4 kb/s Hybrid MELP/CELP Coder 283

9.3.4 Limitations of Existing Hybrid Coders 284

9.4 Synchronized Waveform-Matched Phase Model 285

9.4.1 Extraction of the Pitch Pulse Location 286

9.4.2 Estimation of the Pitch Pulse Shape 292

9.4.3 Synthesis using Generalized Cubic Phase Interpolation 297

9.5 Hybrid Encoder 298

9.5.1 Synchronized Harmonic Excitation 299

9.5.2 Advantages and Disadvantages of SWPM 301

9.5.3 Offset Target Modification 304

9.5.4 Onset Harmonic Memory Initialization 308

9.5.5 White Noise Excitation 309

9.6 Speech Classification 311

9.6.1 Open-Loop Initial Classification 312

9.6.2 Closed-Loop Transition Detection 315

9.6.3 Plosive Detection 318

9.7 Hybrid Decoder 319

9.8 Performance Evaluation 320

9.9 Quantization Issues of Hybrid Coder Parameters 322

9.9.1 Introduction 322

9.9.2 Unvoiced Excitation Quantization 323

9.9.3 Harmonic Excitation Quantization 323

9.9.4 Quantization of ACELP Excitation at Transitions 331

9.10 Variable Bit Rate Coding 331

9.10.1 Transition Quantization with 4 kb/s ACELP 332

9.10.2 Transition Quantization with 6 kb/s ACELP 332

9.10.3 Transition Quantization with 8 kb/s ACELP 333

9.10.4 Comparison 334

9.11 Acoustic Noise and Channel Error Performance 336

9.11.1 Performance under Acoustic Noise 337

9.11.2 Performance under Channel Errors 345

9.11.3 Performance Improvement under Channel Errors 349

9.12 Summary 350

Bibliography 351

10 Voice Activity Detection 357

10.1 Introduction 357

10.2 Standard VAD Methods 360

10.2.1 ITU-T G.729B/G.723.1A VAD 361

10.2.2 ETSI GSM-FR/HR/EFR VAD 361

10.2.3 ETSI AMR VAD 362

10.2.4 TIA/EIA IS-127/733 VAD 363

10.2.5 Performance Comparison of VADs 364

10.3 Likelihood-Ratio-Based VAD 368

10.3.1 Analysis and Improvement of the Likelihood Ratio Method 370

10.3.2 Noise Estimation Based on SLR 373

10.3.3 Comparison 373

10.4 Summary 375

Bibliography 375

11 Speech Enhancement 379

11.1 Introduction 379

11.2 Review of STSA-based Speech Enhancement 381

11.2.1 Spectral Subtraction 382

11.2.2 Maximum-likelihood Spectral Amplitude Estimation 384

11.2.3 Wiener Filtering 385

11.2.4 MMSE Spectral Amplitude Estimation 386

11.2.5 Spectral Estimation Based on the Uncertainty of Speech Presence 387

11.2.6 Comparisons 389

11.2.7 Discussion 392

11.3 Noise Adaptation 402

11.3.1 Hard Decision-based Noise Adaptation 402

11.3.2 Soft Decision-based Noise Adaptation 403

11.3.3 Mixed Decision-based Noise Adaptation 403

11.3.4 Comparisons 404

11.4 Echo Cancellation 406

11.4.1 Digital Echo Canceller Set-up 411

11.4.2 Echo Cancellation Formulation 413

11.4.3 Improved Performance Echo Cancellation 415

11.5 Summary 423

Bibliography 426

Index 429