Adaptive Filters Introduction

Adaptive Filters –Introduction . ... Fundamentals of Adaptive Filtering, ... Speech enhancement (hands-free systems,...

0 downloads 161 Views 3MB Size
Adaptive Filters – Introduction

Gerhard Schmidt Christian-Albrechts-Universität zu Kiel Faculty of Engineering Institute of Electrical and Information Engineering Digital Signal Processing and System Theory

Contents of the Lecture Today:  Boundary conditions

of the lecture

 Contents  Literature hints  Exams

 Notation  Example of an adaptive

Filter  Examples from speech and audio signal processing

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-2

Contents of the Lecture Entire Semester:  Introduction with examples for speech and audio processing  Wiener Filter  Linear Prediction  Algorithms for adaptive filters  LMS und NLMS algorithm  Affine projection  RLS algorithm

 Control of adaptive filters

 Signal processing structures  Applications of linear prediction  Examples for speech and audio processing

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-3

Literature English and German Books: Statistical signal theory:  A. Papoulis: Probability, Random  E. Hänsler: Statistische

Variables, and Stochastic Processes, McGraw Hill, 1965

Signale: Grundlagen und Anwendungen, Springer, 2001

(in German)

Adaptive filters:  E. Hänsler, G. Schmidt: Acoustic  S. Haykin: Adaptive

Echo and Noise Control, Wiley, 2004

Filter Theory, Prentice Hall, 2002

 A. Sayed: Fundamentals

of Adaptive Filtering, Wiley, 2004

Speech processing:  L. R. Rabiner, R. W. Schafer: Digital Processing

of Speech Signals, Prentice Hall, 1978  P. Vary, R. Martin: Digital Speech Transmission, Wiley, 2006  L. R. Rabiner, R. W. Schafer: Introduction to Digital Speech Processing, Now, 2008

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-4

Boundary Contition of the Lecture Credit Points, Exams, Exercises, and Lecture Notes Credit points:  4 ECTS points

Oral exam:  About 30 minutes per student  In the exams period

Exercises:  Two Matlab exercises during the semester

Talks:  Duration about 10 minutes (afterwards short discussion)  Topics will be offered during the lectures (own

suggestions are welcome)

Lecture notes:  Printed versions will be spread at the beginning of each lecture  In the internet via www.dss.tf.uni-kiel.de

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-5

Notation – Part 1 Scalars and Vectors Scalars:

Discrete time index

 Signals:

Coefficient index

 Impulse responses

(time-variant):

 Example for a (real) convolution:

Vectors:

Boldface and lowercase

 Signal vectors:  Impulse response

vectors (time-variant) :

 Example for a real convolution:

Matrices:

Boldface and uppercase Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-6

Notation – Part 2 Random Processes Random variables and processes:  Notation:

No differences between deterministic signals and random processes – different writing styles:  Probability density function:  Stationary random processes:

 Expected values of stationary random processes:

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-7

Notation – Part 3 Correlation

Auto and cross correlation for real, stationary random processes:  Auto-correlation function:

 Cross-correlation

function:

 (Auto) power spectral density:

 (Cross)

power spectral density:

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-8

Notation – Part 4 White Noise Stationary white noise:  Auto-correlation function:

 Auto power spectral density:

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-9

A First Example of an Adaptive Filter – Part 1 Basic Structure

Unknown impulse response

Adaptive filter

+ +

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Local signals

Unknown system Slide I-10

A First Example of an Adaptive Filter – Part 2 Matlab Demo

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-11

Applications of Adaptive Filters Selected Application Areas  Speech coding (e.g. GSM, UMTS)

 Speech enhancement (hands-free systems, hearing aids, public address systems)  Equalization (sending antennas, radar, loudspeakers)  Anti-noise systems (cars and airplanes)  Multi-channel signal processing (beamforming,

submarine localization, layer of earth analysis)

 Missile control  Medical applications (fetal heart rate monitoring, dialysis)  Processing of video signals (cancellation of distortions, image analysis)

 Antenna arrays

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-12

Basis Structures of Adaptive Filters – Part 1 System Identification

Unknown system

+ +

Adaptive filter

Examples:  Line echo cancellation  Cancellation of acoustical echoes

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-13

Basis Structures of Adaptive Filters – Part 2 Inverse Modelling

Unknown system

Adaptive filter

+ Delay Distortions are not depicted!

Examples:  Equalization of amplifiers of transmission antennas  Loudspeaker equalization

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-14

Basis Structures of Adaptive Filters – Part 3 Prediction

Delay

Adaptive filter

+

Examples:  Speech coding in the GSM and UMTS networks  Suppression of carrier signals after demodulation

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-15

Basis Structures of Adaptive Filters – Part 4 Cancellation of Undesired Signals

+

Adaptive filter

Example:  Automotive speech signal enhancement via cancellation

of engine harmonics

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-16

Examples from Speech and Audio Processing Contents Part 1: Automotive hands-free telephone systems  Basics  Solutions  Examples

Part 2: In-car communication systems  Basics  Solutions

 Examples

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-17

Examples from Speech and Audio Processing Part 1

Automotive Hands-Free Telephone Systems

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-18

Automotive Hands-Free Telephone Systems Basics – Electro-Acoustic Transducers Microphones:  Integrated in the rear-view

mirror (example)

 Up to four microphones

Loudspeakers: Loudspeakers of the car stereo (head unit)  coupling > 0 dB  Volume adjustable by the passengers  Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-19

Automotive Hands-Free Telephone Systems Basics – Loudspeaker Enclosure Microphone (LEM) Systems – Part 1 Signal of the remote communication partner

: Excitation signal

: Echo (desired) signal : Local speech signal : Background noise

: Microphone signal

Microphone signal

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-20

Automotive Hands-Free Telephone Systems Basics – Loudspeaker Enclosure Microphone (LEM) Systems – Part 2

Assumption: FIR filter

The loudspeaker enclosure microphone system (LEM system) can be modeled as a linear system with finite memory.

+

+

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-21

Automotive Hands-Free Telephone Systems Basics – Loudspeaker Enclosure Microphone (LEM) Systems – Part 3

0.4

Boundary conditions:  Volume

of a passenger compartment: 5 … 15 m³

0.3 0.2 0.1

Properties:  Short delay

0

 Direct -0.1

 Early reflections  Diffuse sound (decays

-0.2

logarithmically in amplitude)

-0.3 -0.4 0

sound after 3 … 4 ms

5

10

15

20 Time in ms

25

30

35

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

40

Slide I-22

Automotive Hands-Free Telephone Systems Basics – Background Noise and its Components

External components:  Engine noise

 Wind noise  Tire noise

Internal components:  Air conditioning  Defrost

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-23

Automotive Hands-Free Telephone Systems A Basic System With Two Adaptive Filters

Echo cancellation filter

Noise suppression filter

+

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-24

Automotive Hands-Free Telephone Systems An Adaptive Filter for Cancellation of Acoustical Echoes Loudspeaker

enclosure microphone system FIR model

Adaptive echo cancellation filter

+

(system parameters are unknown, only input and output signals are measurable)

+

+

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-25

Automotive Hands-Free Telephone Systems Maximal Achievable Echo Reduction – Part 1 Derivation during the lecture …

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-26

Automotive Hands-Free Telephone Systems Maximal Achievable Echo Reduction – Part 2 5

Boundary conditions: Maximum echo attenuation in relation to the filter length

0

 White noise as

excitation signal

-5

 Ideal convergence,

-10

meaning that all filter coefficients of the adaptive filter are equal to the corresponding ones of the impulse response.

dB

-15 -20 -25 -30

 Linear loudspeakers, -35 -40

microphones, and amplifiers 0

50

100

150

200

250

300

350

400

450

Filter length

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-27

Automotive Hands-Free Telephone Systems A Basic System With Two Adaptive Filters

Echo cancellation filter

Noise suppression filter

+

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-28

Automotive Hands-Free Telephone Systems Residual Echo and Noise Suppression Remaining echoes....

... and local background noise

+

Local speech signal

Approach according to Wiener (next lecture):

Cross power spectral density of the distorted input signal and the desired output signal

Auto power spectral density of the distorted input signal Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-29

Automotive Hands-Free Telephone Systems A Basic System With Two Adaptive Filters – Audio Examples

Stereo signals (16 kHz): Left:

Right:

Received signal ...

Sent signal ...

... of the remote communication partner

Initial filter convergence: Adaptation at the beginning of the call

Transmission to the communication partner (channel delay: about 180 ms)

Received signal („Hearing channel“ of the remote communication partner)

Double talk:

Remote communication partner

Enclosure dislocations:

Both partners speak simultaneously

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Without Wiener filter With Wiener filter Slide I-30

Automotive Hands-Free Telephone Systems Enhanced Systems

Improvements:  Improved noise suppression

by adaptive combination of several microphone signals (beamforming)

 Further

improvements by applying adaptive filters for different kinds of distortions

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-31

Automotive Hands-Free Telephone Systems Microphone Array Using Four Sensors (Integrated into the Rear-View Mirror)

Rear-view mirror

Microphone

module



Cheap realization by means of an integrated microphone module.



A fixed steering direction can be used for the driver – the steering angle varies only in a small range (62° - 75°).



The array can be used for the driver and for the passenger simultaneously.



Cardioid microphones are usually applied (± 3 dB sensitivity).

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-32

Automotive Hands-Free Telephone Systems Beamforming – Introduction

Desired signal

Distortion Adaptive filters

Beamformer:  Minimizing the output power with respect to one or more constraints (signals from a desired direction must pass the structure without distortion)  The desired direction is known in automotive applications (at least approximately)  The performance of adaptive filtering is limited by sensor tolerances and multipath propagation within the passenger compartment Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-33

Automotive Hands-Free Telephone Systems Beamforming – Adaptive Structure Output of the so-called generalized sidelobe canceller

Summation path

+

+ Delay

+

Adaptive filter

Blocking path

„Griffith-Jim“ beamformer (generalized sidelobe canceller)

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-34

Automotive Hands-Free Telephone Systems Beamforming – Audio Examples 

4-channel beamformer



Loudspeaker on the passengers seat (undesired signal)



Adaptive filtering of the microphone signal results in an SNR improvement of about 15 dB.

Single microphone Fixed beamformer Adaptive beamformer

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-35

Automotive Hands-Free Telephone Systems Beamforming – Performance of Speech Recognition Systems Automotive wind, engine, tire noise Basic commands (120and km/h)

Noise a defroster Basic produced commands by (defrost on)

With permission from Eberhard Hänsler, Gerhard Schmidt, Acoustic Echo and Noise Control, Wiley, 2004

Speech and noise were mixed artificially to obtain different signal-to-noise ratios.  About 30 command words for controlling the radio and phone system were used.  16 subjects (9 male, 7 female) participated in the test. 

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-36

Automotive Hands-Free Telephone Systems Involved Signal Processing Units – Start

Microphone array

Acoustic coupling from the loudspeaker to the microphone(s)

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Telephone or speech dialog system

Slide I-37

Automotive Hands-Free Telephone Systems Involved Signal Processing Units – Bandwidth Extension Bandwidth extension Missing frequency components were estimated and resynthesized. Effect: The speech quality (not the intelligibility) of the received signal is improved. Microphone array

Acoustic coupling from the loudspeaker to the microphone(s)

Telephone or speech dialog system

Bandwidth extension

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-38

Automotive Hands-Free Telephone Systems Involved Signal Processing Units – Automatic Gain and Equalization Adjustment Volume and equalization control The (broadband) playback volume is adjusted automatically with respect to the noise measured in the car. In addition also the spectrum can be shaped in order to improve the perceived signal quality. Microphone array

Acoustic coupling from the loudspeaker to the microphone(s)

Telephone or speech dialog system

Adaptive volume and equalization control

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Bandwidth extension

Slide I-39

Automotive Hands-Free Telephone Systems Involved Signal Processing Units – Adaptive Limiter Adaptive limiter Adaptive adjustment of the parameters of a limiter in order to avoid microphone clipping by those loudspeakers that are close to the microphones (e.g. so-called center speaker). Microphone array

Acoustic coupling from the loudspeaker to the microphone(s)

Telephone or speech dialog system

Adaptive limiter

Adaptive volume and equalization control

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Bandwidth extension

Slide I-40

Automotive Hands-Free Telephone Systems Involved Signal Processing Units – Echo Cancellation Echo cancellation The signals emitted by the loudspeakers are reflected by windows, etc. These reflected signals as well as directly coupled signals are also recorded by the microphones. Microphone array

Acoustic coupling from the loudspeaker to the microphone(s)

To decouple the electro-acoustic system, the echo signals are estimated and subtracted from the microphone signal.

Telephone or speech dialog system

Echo cancellation

Adaptive limiter

Adaptive volume and equalization control

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Bandwidth extension

Slide I-41

Automotive Hands-Free Telephone Systems Involved Signal Processing Units – Beamforming Beamforming The microphone signals are filtered such that a predefined direction is kept open, while other directions are attenuated as much as possible. Effect: Directional distortions can be suppressed. Microphone array

Acoustic coupling from the loudspeaker to the microphone(s)

Beamforming

Telephone or speech dialog system

Echo cancellation

Adaptive limiter

Adaptive volume and equalization control

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Bandwidth extension

Slide I-42

Automotive Hands-Free Telephone Systems Involved Signal Processing Units – Noise and Residual Echo Suppression Background noise and residual echo suppression Despite beamforming and echo cancellation several remaining undesired signal components are still audible. Effect: Stationary background noise and residual echoes can be suppressed. Microphone array

Acoustic coupling from the loudspeaker to the microphone(s)

Noise and Beamforming echo suppression

Telephone or speech dialog system

Echo cancellation

Adaptive limiter

Adaptive volume and equalization control

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Bandwidth extension

Slide I-43

Automotive Hands-Free Telephone Systems Involved Signal Processing Units – Wind Buffet Removal Wind buffet suppression Open windows and defrost on might cause wind buffets. Effect: A detection optimized for those undesired signals finds wind buffets and replaces the signal with so-called comfort noise. Microphone array

Acoustic coupling from the loudspeaker to the microphone(s)

Noise and Wind buffet Beamforming echo suppression removal

Telephone or speech dialog system

Echo cancellation

Adaptive limiter

Adaptive volume and equalization control

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Bandwidth extension

Slide I-44

Automotive Hands-Free Telephone Systems Involved Signal Processing Units – Removal of “Transients” Suppression of transients Transient signal, such as the noise of an indicator or a wind shield wiper, cause problems for voice recognitions signals (voice activity detection). Effect: Short impulsive distortions are suppressed. Microphone array

Acoustic coupling from the loudspeaker to the microphone(s)

Noise and Wind buffet Beamforming echo suppression removal

Suppression of transients

Telephone or speech dialog system

Echo cancellation

Adaptive limiter

Adaptive volume and equalization control

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Bandwidth extension

Slide I-45

Automotive Hands-Free Telephone Systems Involved Signal Processing Units – Adaptive Equalization Adaptive equalization For compensation of different microphone-speaker distances and room characteristics, a (blind) equalization can be performed adaptively. Effect: The signal sounds more natural. Microphone array

Acoustic coupling from the loudspeaker to the microphone(s)

Noise and Wind buffet Beamforming echo suppression removal

Suppression of transients

Adaptive equalization

Telephone or speech dialog system

Echo cancellation

Adaptive limiter

Adaptive volume and equalization control

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Bandwidth extension

Slide I-46

Automotive Hands-Free Telephone Systems Involved Signal Processing Units – Summary Bandwidth extension

Volume and equalization control

Adaptive limiter

Missing frequency components were estimated and resynthesized. Effect: The speech quality (not the intelligibility) is improved.

The (broadband) playback volume is adjusted automatically with respect to the noise measured in the car. In addition also the spectrum can be shaped in order to improve the perceived signal quality.

Adaptive adjustment of the parameters of a limiter in order to avoid microphone clipping by those loudspeakers that are close to the microphones (e.g. so-called center speaker).

Echo cancellation

Beamforming

Noise and residual echo suppression

To decouple the electro-acoustic system, the echo signals are estimated and subtracted from the microphone signal.

The microphone signals are filtered such that a predefined direction is kept open, while other directions are attenuated.

Despite beamforming and echo cancellation several remaining undesired signal components are still audible.

Effect: Directional distortions can be suppressed.

Effect: Stationary background noise and residual echoes can be suppressed.

Wind buffet suppression

Suppression of transients

Adaptive equalization

Open windows and defrost on cause might cause wind buffets.

Transient signal, such as the noise of an indicator or a wind shield wiper, cause problems for voice recognitions signals.

For compensation of different microphonespeaker distances and room characteristics, a (blind) equalization can be performed adaptively.

Effect: A detection optimized for those signals finds wind buffets and replaces the signal with so-called comfort noise.

Effect: Short impulsive distortions are suppressed.

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Effect: The signal sounds more natural.

Slide I-47

Examples from Speech and Audio Processing Part 2

In-Car Communication Systems

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-48

In-Car Communication Systems Motivation Passenger compartment

Current situation:  Communication between

passengers is difficult, because of the acoustic loss (especially front to rear).

 Driver turns around – road safety is reduced.

Driving direction

-5 … -15 dB*

 Front

passengers have to speak louder than normal – longer conversations will be tiring.

Solutions:  Improve

the speech quality and intelligibility by means of an intercom system.

*Acoustic loss (referred to the ear of the driver)

Application: 

Mid and high-class automobiles, which are already equipped with the necessary audio and signal processing devices.



Vans, etc. – systems with reduced complexity.

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-49

In-Car Communication Systems Algorithmic Overview

the speech quality and intelligibility by means of an ICC system.  The ICC system records the speech by means of microphones and improves the communication by playing back the signals via those loudspeakers that are close to the listening passengers.

Loudspeakers

Front passenger

Solution:  Improve

Rear passengers Microphones Driver Loudspeakers Passenger compartment Feedback and noise suppression

Mixer

Adaptive splitter, Automatic equalizer, delay, gain control, noise dependent limiter gain adjustment

Clipping detection, highpass filtering, speaker localization, beamforming

Feedback cancellation

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-50

In-Car Communication Systems Results of a Comparison Mean Opinion Score (CMOS) Test 0 km/h, car parked close to a motorway  19.7 % prefer the system

to be switched off  29.7 % have no preference

 50.6 % prefer an activated

system

130 km/h, on a motorway 

4.3 % prefer the system to be switched off



7.1 % have no preference

 88.6 % prefer an activated

system With permission from Eberhard Hänsler, Gerhard Schmidt (eds.), Topics in Acoustic Echo and Noise Control, Springer, 2006

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-51

In-Car Communication Systems Diagnostic Rhyme Tests (DRT) and Modified Rhyme Tests (MRT) On a parking area beside motorway (0 km/h):  No significant difference (95.2 system off versus 95.0 % system on).  Due to the automatic gain adjustment the intercom system operates with

only very small gain at these noise levels.

On a motorway (130 km/h):  Significant improvement

of the DRT error rate.  Nearly 50 % error

reduction

(85.4 % correct answers increased to 92.2 % correct answers).

With permission from Eberhard Hänsler, Gerhard Schmidt (eds.), Topics in Acoustic Echo and Noise Control, Springer, 2006

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-52

Adaptive Filters – Introduction Summary and Outlook

This week: Boundary conditions of the lecture  Contents  Literature hints  Exams  Notation  Example of an adaptive Filter  Examples from speech and audio signal processing

Next week:  Wiener filter  Noise suppression

Digital Signal Processing and System Theory| Adaptive Filters | Introduction

Slide I-53