Débruitage multicapteur appliqué à la téléphonie mains-libres en automobile

HAL Id: tel-01156542

https://pastel.archives-ouvertes.fr/tel-01156542

Submitted on 27 May 2015

HAL is a multi-disciplinary open access

archive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Charles Fox

To cite this version:

Charles Fox. Débruitage multicapteur appliqué à la téléphonie mains-libres en automobile. Traitement du signal et de l’image [eess.SP]. Télécom ParisTech, 2013. Français. �NNT : 2013ENST0074�. �tel-01156542�

beamfor ming Minimum Variance Distortionless Response

Speech Distor tion Weighted Multichannel Wiener Filter

2 Env i r onnem ent acoust i que 29

3 B eam f or m i ng M i ni m um V ar i ance D i st or t i onless Response ( M V D R ) adapt

a-t i f 61

4 A nnul at i on de br ui t adapt at ive r ési st ant e aux fuit es de par ol e ( C R -A N C ) 87

C oncl usi on 135

A nnex es 139

A M at r i ce sp ect r al e et cohér ence 139

B Least -M ean Squar e ( L M S) dans l e dom ai ne fr équent i el 143 C Si mul at i on de sal l e r éver b ér ant e par m ét hode d’ i m age 147

R éfér ences 151

A EC Acoustic Echo Cancel lation

A N C Adaptive Noise Cancel lation

A PA Affine Projection Algorithm

C R -A N C Crosstalk Resistant

Adaptive Noise Cancel lation D SP

EC M Electret Condenser Microphone EQM

F B F Fixed Beamfor ming

F F T Fast Fourier Transform

G SC Generalized Sidelobe Cancel ler

G SV D Generalized Singular Value Decomposition

H M M Hidden Mar kov Model

I T U -T I nternational Telecommunication Union - Telecommunication Standardization Sector

K LT Kar hunen-Loeve Transform

L M S Least-Mean Square L SA Log-Spectral Amplitude M OS Mean Opinion Score M SC Mean-Squared Coherence

M V D R Minimum Var iance Distor tionless Response

M W F Multichannel Wiener Filter

N L M S Normalized LMS

N M F Nonnegative Matrix Factor ization

N R Noise Reduction

OL A Overlap-Add

OM -L SA Optimal ly Modified Log-Spectral Amplitude P ESQ Perceptual Evaluation of Speech Quality P C M Pulse Code Modulation

SI M O Single I nput Multiple Output

SP P Speech Presence Probability

SSL T F D T F D i T FC T

E T ∗ H x t x t X f x t X f x t ak k a φx f x t φx y f x t y t Rx τ x t Rx τ x t x f x t x x ⊗ δt t cs − 1 x x ⌊x⌋ x x x Ik× k k × k 0k× l k × l

Generalized Sidelobe Cancel ler

Adaptive Noise Cancel lation

Crosstalk Resis-tant Adaptive Noise Cancel lation

◦

◦

Mean-Squared Coherence

Endfire Broadside

Broadside

Broadside Endfire

R f f Fs d cs Endfire |Hs f − HA N C f |2 | h|2 |Hs f − Hbf |2 Voice Activity Detector Endfire Multichannel Wiener Filter R f ρ R f ρ R f ρ

ρ ρ

Wideband

Perceptual Evaluation of Speech Quality

a posterior i

F igur e

L e br ui t am biant

Bruit

_{Habit acle}

F igur e

Acoustic Echo Cancel lation Noise Reduction

beamforming a prior i

beamforming Minimum Variance Distortionless Response

Adaptive Noise Cancel lation

beamforming

Speech Distortion Weighted Multichannel Wiener Filter

8

16

19

25

Speech Distortion Weighted Multichannel Wiener Filter

A i des audi t i ves

T él écom muni cat i ons

I nt er faces hom m e-m achi ne

A nal y se d’ enr egi st r em ent s a posteriori

1.1. Généralit és sur le débruit age de parole

M ét hodes t em p or el l es

F i l t r age de W i ener [W i ener , 1950]

F i l t r age de W i ener p ondér é [F l or enci o and M alvar , 2001]

F i l t r e de K al m an [G annot et al ., 1998]

Hidden Markov Model

F igur e

Karhunen-Loeve Transform

K LT [Ephr ai m and V an Tr ees, 1995]

Generalized Singular Value Decomposition

M ét hodes par t r ansfor m ée en ondel et t es [Seok and B ae, 1997]

M ét hodes fr équent i el l es

Ex t ension des m ét hodes t em p or el l es

1.1. Généralit és sur le débruit age de parole

Est i m at i on d’ am pl i t ude sp ect r al e [Ephr ai m and M al ah, 1984]

Est i m at i on de l a l og-am pl i t ude sp ect r al e [Ephr ai m and M al ah, 1985]

M ét hodes par fact or i sat i on de m at r i ces

Non-negative Matr ix Factorization

F igur e

Ex t ension des m ét hodes m ono-capt eur

F i l t r age de W iener

M ét hodes de sous-espace

Fact or i sat i on des m ét hodes fr équent i el les [B al an and R osca, 2002]

beamforming Minimum Variance Distortionless Response

1.1. Généralit és sur le débruit age de parole

M ét hodes pur em ent m ult i -capt eur s

B eam for m i ng

matched beamforming

Generalized Sidelobe Cancel ler

F igur e

Fixed Beamfor ming

F igur e Adaptive Noise Cancel lation

1.1. Généralit és sur le débruit age de parole

Figur e

Speech Presence Probability

1.2. Cont ext e de l’ét ude : t éléphonie mains-libres en aut omobile

F igur e

a priori

beamforming

Signal

prÈ-dÈbruit È

Et age

m ult

i-capt eur

DÈbruitage

m ono

capt eur

{

Objet de

la t hËse

Exist ant

{

F igur e m xm t m∈ M s t s t { gT ⊗x } t

1.3. Débruit age fréquent iel

x t x1 t xM t T g t g1 t gM t T

Multichannel Wiener Filter

s t b t t x t s t b t Overlap-Add F igur e

f n

Xn f Sn f Bn f

n

S f B f

φs f φb f

Est i m at i on sépar ée de l a phase et de l ’ am pl i t ude sp ect r al e

S f

S f A f eiψ(f ) A f > ψf ∈ π

A f ψ f

[S f [A f ei [ψ(f )

Est i m at eur de phase

[ ψ f X f

Est i m at eur basé sur l a l og-am pl i t ude sp ect r ale

E E h | [A f − A f |2| X f i X f [A f eE[ln A (f )| X (f )] a priori a posteriori

1.3. Débruit age fréquent iel RSBpr i o f φs f φbf RSBpost f |X f |2 φb f v f v f RSBpr i o f RSBpr i o f RSBpost f [A f RSBpr i o f RSBpr i o f Z _∞ v(f ) e− t t dt ! | { z } GL S A |X f |

a prior i a posterior i

Log-Spectral Amplitude [S f GL SA f X f R em ar que 1: φs f φbf φb f φs f φs f φx f − φb f φx f M odèl e M m M xm t xm t { hm ⊗s} t bm t hm t m s t bm t m

Figur e Xm f Hm f S f Bm f X f H f S f B f S f B f φs f S f bf b t H f M

Est i m at i on sépar ée de phase et d’ am pl i t ude sp ect r al e

S f A f eiψ(f )

Est i m at eur de phase

[ ψ f E ψf | X f E ψf | X f E ψf | T X f T X f H f H bf − 1X f H f H b f − 1H f [ ψf T X f

1.3. Débruit age fréquent iel

X

Est i m at eur d’ am pl i t ude sp ect r al e

[A f eE[ln A (f )| X (f )] eE[ln A (f )| X (f )] eE[ln A (f )| T (X (f ))] [S f RSBpr i o f RSBpr i o f Z _∞ v(f ) e− t t dt ! | { z } P ost − F i l t r e: GL S A H f H bf − 1X f H f H bf − 1H f RSBpr i o f φs f H f H b f − 1H f RSBpost f |T X f |2H f H bf − 1H f v f RSBpr i o f RSBpr i o f RSBpost f T X f H f H b f − 1H f − 1 R em ar que 2: \W f G E h |GHX f |2 i GHH f \W f Σb(f )− 1H (f ) H (f )H_Σ b(f )− 1H (f ) T X f beamforming beamforming

F igur e

Hf₁ X f S f B f

Hf₀ X f B f

Hf₁ f Hf₀

f

P r obabi l i t é condi t i onnel l e de pr ésence de par ol e

p Hf₁|X f p Hf₁|X f " p Hf₀ − p Hf₀ RSBpr i o f e − R SBpost(f )_{1+ R S B pr i o ( f )}R S B pr i o ( f ) #− 1 p Hf₀ a pr iori

M odi fi cat i on de l ’ est i m at eur de l og-am pl i t ude sp ect r al e

ρ S f E ρS f | X f p Hf₀ E h ρS f | X f Hf₀ i p Hf₁ E h ρS f | X f Hf₁ i eE[ln |S(f )|| X (f )] ep(Hf0)E ln |S(f )|| X (f ) H f 0 _ep(H f 1)E ln |S(f )|| X (f ) H f 1

1.4. Ét age mult i-capt eurs Gm i n eE ln S(f )| X (f ) Hf0 eE ln |S(f )|| X (f ) Hf0 _G m i n|X f | eE[ln |S(f )|| X (f )] eE ln |S(f )|| X (f ) Hf1 p G1− p_{m i n} GL SA f pG1− pm i n|X f | GL SA f p p H1f|X f Optimal ly

Modified Log-Spectral Amplitude

L i m i t at ions du G SC H

x1 t s t b t x2 t { hb⊗b} t hbt x1 t s t b t x2 t { hs⊗s} t { hb⊗b} t hs t x1 t b t x2 t { hb⊗b} t hb

Voice Activity Detector

hb [S f X1 f − X2 f Hbf S f − Hs f Hb f

Parole dist ordue

Annulat ion de bruit

VAD = 1

VAD = 0

1.4. Ét age mult i-capt eurs Hout Hout f E h S f B f S f − Hs(f ) Hb(f ) i E h |S f − Hs(f ) Hb(f ) | 2i − Hs f Hb f − 1 S f xm t { hm ⊗s} t bm t sm t { hm⊗s} t m L w M L w g E |b1 t − g H_{x t |}2 x t x1 t H x2 t H xM t H H xm t xm t − L xm t H sm t bm t s t b t xm L × x t M L × b s t x1 t − wHx t w g h E |b1 t − gHb t |2 i | { z } eb h E |gHs t |2 i | { z } es es eb wρ g E h |b1 t − gHb t |2 i ρE h |gHs t |2 i

wρ ρE s t s t H E b t b t H − 1 E b t b∗₁ t ρ ρ ρ w δ t x1 t E s t s t H E b t b t H beamforming

30 36 Wideband 42 50 58

F igur e

2.1. Syst ème d’acquisit ion C ondi t i onnem ent

Electret Condenser Microphone

C ar t e N at i onal I nst r um ent s

Pulse Code Modulation

Or di nat eur 1

F igur e ◦ Tabl e −15 −15 −10 −10 −5 −5 0 0 5 dB 5 dB 0o 30o 60o 90o 120o 150o 180o 210o 240o 270o 300o 330o 125 Hz 250 Hz 500 Hz 1 kHz 2 kHz 4 kHz F igur e ◦

2.1. Syst ème d’acquisit ion

0.2 0.4 0.6 0.8 30° 210° 240° 270° 300° 150° 330° 180° 0° F igur e Tabl e −15 −15 −10 −10 −5 −5 0 0 5 dB 5 dB 0o 30o 60o 90o 120o 150o 180o 210o 240o 270o 300o 330o 125 Hz 250 Hz 500 Hz 1 kHz 2 kHz 4 kHz F igur e ◦ ∼ ∼ ◦

2.1. Syst ème d’acquisit ion

F igur e

Om ni di r ect i onnel

0 1000 2000 3000 4000 5000 6000 7000 8000 −10 −5 0 5 10 15 20 25 Fréquence (Hz) R S B e n e n tr é e ( d B ) omnidirectionnel unidirectionnel F igur e Mean-Squared Coherence x t y t xy f |φx y f |2 φx f φy f

2.2. Sources acoust iques en environnement aut omobile φx f 6 φy f 6 φab f a t b t φ_ab f Z _+∞ −∞ E h a t b t − τ i e− 2iπfτdτ φx f x t x y x t { h⊗y} t v t v t y t x y f |H f |2|φy f |2 |H f |2_φ y f φv f φy f x t { h⊗_{y} t x t y t} " x t y t # x y Est i m at i on de l a M SC a t b t N n a b f An f Bn f d φab f N N − 1_X n= 0 An f B ∗ n f \M SC f | P _{N − 1} n = 0 An f B ∗ n f |2 P _{N − 1} n = 0 An f A∗n f P _{N − 1} n = 0 Bn f Bn∗ f φx f φy f ∼

C ondi t i ons d’ enr egi st r em ent

2.2. Sources acoust iques en environnement aut omobile Temps (s) F ré q u e n c e ( H z) 1 2 3 4 5 6 7 0 1000 2000 3000 4000 5000 6000 7000 8000 −60 −50 −40 −30 −20 −10 0 10 F igur e Temps (s) F ré q u e n c e ( H z ) 1 2 3 4 0 1000 2000 3000 4000 5000 6000 7000 8000 −60 −50 −40 −30 −20 −10 0 10 0 1000 2000 3000 4000 5000 6000 7000 8000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fréquence (Hz) M S C Figur e

Temps (s) F ré q u e n c e ( H z ) 0.5 1 1.5 2 0 1000 2000 3000 4000 5000 6000 7000 8000 −60 −50 −40 −30 −20 −10 0 10 0 1000 2000 3000 4000 5000 6000 7000 8000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fréquence (Hz) M S C F igur e

2.2. Sources acoust iques en environnement aut omobile Temps (s) F ré q u e n c e ( H z ) 0.5 1 1.5 2 2.5 3 3.5 0 1000 2000 3000 4000 5000 6000 7000 8000 −60 −50 −40 −30 −20 −10 0 10 0 1000 2000 3000 4000 5000 6000 7000 8000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fréquence (Hz) M S C Figur e W i deban d

I nternational Telecommunication Union - Telecommunication Standardization Sec-tor

Wideband

C apt eur s om ni di r ect i onnels

f df cs 2 d cs f x πx πx

M SC δ δ δ x1x2 y1y2 δ δ3 − δ δ2 z1z2 δ δ δ δ2 − δ δ3 i z1 z2 δ δ − δ δ2 2 F igur e 0 1000 2000 3000 4000 5000 6000 7000 8000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fréquence (Hz) M S C (a) (b) 0 1000 2000 3000 4000 5000 6000 7000 8000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fréquence (Hz) M S C (c) (d) (e) F igur e

2.3. Champ de bruit

A nt enne m i x t e om ni di r ect i onnel - car di oïde

M SC δ δ δ i θ δ2 δ δ − δ 2 θ F igur e 0 1000 2000 3000 4000 5000 6000 7000 8000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Fréquence (Hz) M S C (a) (b) F igur e

cs − 1 0 1000 2000 3000 4000 5000 6000 7000 8000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fréquence (Hz) M S C 4cm 8cm 12cm 0 1000 2000 3000 4000 5000 6000 7000 8000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fréquence (Hz) M S C 4cm 8cm 12cm F igur e 0 1000 2000 3000 4000 5000 6000 7000 8000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fréquence (Hz) M S C 4cm 8cm 12cm 0 1000 2000 3000 4000 5000 6000 7000 8000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fréquence (Hz) M S C 4cm 8cm 12cm F igur e

2.3. Champ de bruit 0 1000 2000 3000 4000 5000 6000 7000 8000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Fréquence (Hz) M S C 4cm 8cm 12cm 0 1000 2000 3000 4000 5000 6000 7000 8000 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Fréquence (Hz) M S C 4cm 8cm 12cm F igur e C as du br ui t t r ès cohér ent xr ef x1 N h g |xH₁ − gHXr ef|2 x1 x1 x1 N − L H Xr ef

Xr ef xr ef xr ef N − L xr ef L − xr ef N − L h Xr efXHr ef − 1 Xr ef x1

Co e

c i e n t u l t r e

F igur e

2.3. Champ de bruit −30 −20 −10 0 10 20 30 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 Temps (ms) V a le u r 4cm 8cm 12cm −30 −20 −10 0 10 20 30 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 Temps (ms) V a le u r 4cm 8cm 12cm F igur e 0 1000 2000 3000 4000 5000 6000 7000 8000 −50 −40 −30 −20 −10 0 10 20 30 Fréquence (Hz) D S P ( d B ) F igur e

2.4. Propagat ion de la voix 0 6.3 12.5 18.7 25 31.2 37.5 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 Temps (ms) V a le u r F igur e 0 1000 2000 3000 4000 5000 6000 7000 8000 −40 −30 −20 −10 0 10 20 Fréquence (Hz) A m p li tu d e ( d B ) F igur e

Figur e Endfire Broadside 0 1000 2000 3000 4000 5000 6000 7000 8000 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fréquence (Hz) M S C 4cm 8cm 12cm F igur e

2.4. Propagat ion de la voix 0 1000 2000 3000 4000 5000 6000 7000 8000 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fréquence (Hz) M S C 4cm 8cm 12cm F igur e 0 1000 2000 3000 4000 5000 6000 7000 8000 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Fréquence (Hz) M S C 4cm 8cm 12cm F igur e N N K × L K L k ∈ _{K −}

\ M SCk f | _{n = kL} An f Bn f |2 P (k+ 1)L − 1 n = kL An f A ∗ n f P (k+ 1)L − 1 n= kL Bn f B ∗ n f σ2 M SC f _{K −} K − 1_X k= 0 \ M SCk f − \M SC f 2 \M SC f_q N σ2 M SC f N L K 0 1000 2000 3000 4000 5000 6000 7000 −40 −30 −20 Fréquence (Hz) D S P (d B ) 0 1000 2000 3000 4000 5000 6000 7000 0 0.5 1 Fréquence (Hz) M S C 0 1000 2000 3000 4000 5000 6000 7000 0 0.05 0.1 Fréquence (Hz) D é v ia ti o n S ta n d a rd F igur e

2.4. Propagat ion de la voix T ( ) F igur e Broadside Broadside Endfire C as B r oadsi de Broadside

−2.5 −1.9 −1.3 −0.62 0 0.62 1.3 1.9 2.5 −0.2 0 0.2 0.4 0.6 0.8 1 Temps (ms) V a le u r 4cm 8cm 12cm −2.5 −1.9 −1.3 −0.62 0 0.62 1.3 1.9 2.5 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Temps (ms) V a le u r 4cm 8cm 12cm −2.5 −1.9 −1.3 −0.62 0 0.62 1.3 1.9 2.5 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Temps (ms) V a le u r 4cm 8cm 12cm F igur e Broadside C as Endfi r e Endfire

2.4. Propagat ion de la voix −2.5 −1.9 −1.3 −0.62 0 0.62 1.3 1.9 2.5 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 Temps (ms) V a le u r 4cm 8cm 12cm −2.5 −1.9 −1.3 −0.62 0 0.62 1.3 1.9 2.5 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Temps (ms) V a le u r 4cm 8cm 12cm −2.5 −1.9 −1.3 −0.62 0 0.62 1.3 1.9 2.5 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 Temps (ms) V a le u r 4cm 8cm 12cm F igur e Endfire Endfire

Sur l e br ui t

Sur l a par ol e

B eam f or m i n g M i n i m um V ar i an ce

D i st or t i on less Respon se

62 64 78 79 85 beamforming

beamforming

M

Single I nput Multiple Output m t xm t { hm ⊗s} t bm t s t hm t bm t s hm hm m∈_{[1 M ]} x t { h⊗_{s} t b t} x t x1 t xM t h t b t r 1₂ m fk n Xm n fk √ N ⌊αn N_X⌋+ N t =⌊αn N⌋+ 1 xm t e− 2iπfkt

3.1. Rappels sur le modèle et les not at ions N α − r r ∈ fk k N k N − fk Fs Fs n f Xm n fk Hm fk Sn fk Bm n fk hm M Xn fk H fk Sn fk Bn fk Xn fk X1 n fk XM n fk T n b t Rbτ τ Rbτ E h b t τ b t H i bf X τ ∈ZZZ Rb τ e− 2iπfτ s t Rs τ E s t τ s t ∗ φs f X τ ∈ZZZ Rs τ e− 2iπfτ φa f a t φab f a t b t

Log-Spectral Amplitude

beam-for ming

\

T Xn fk HH fk bfk − 1Xn fk HH fk b fk − 1H fk H1 fk { h1⊗s} t T X fk H fk b fk b beamfor ming H f

3.2. Est imat ion de la propagat ion relat ive de la parole Figur e H fk f ∈ H fk αe− 2iπfkFs∆d cs α d2 d d 2 d cs H f 2

2. Nous avons ut ilisé le logiciel Room Impulse Response Generat or de E.H.P Habet s, disponible à l’adresse

Figur e φr ever b P _+∞ t = −∞ |hsal l e t − hdi r ect t |2 P _+∞ t = −∞ |hdi r ect t |2 hsal l e hdi r ect s t b t b b t { h ⊗s} t B f [ H f H b f − 1B f [ H f H bf − 1H f[ S f [ H f H bf − 1H f S f [ H f H b f − 1H f[ [ H f H f bf

3.2. Est imat ion de la propagat ion relat ive de la parole b f " φb f φbf # φb φb1 f φb2 f φbf RSBs E |s t |2 E h |b t |2 i s t b t RSBe E |s t |2 E |b1 t |2 RSB RSBs RSBe D E |s t − s t | 2 E |s t |2

10 0 10 20 12 10 8 6 4 2 0 2 4 6 1 1.5 2 2.5 3 Energie relative de la reÈverbeÈration (dB) Gain en RSB (dB) RSB díentreÈe (dB) 5 0 5 10 15 20 15 10 5 0 5 16 14 12 10 8 Energie relative de la

reÈverbeÈration (dB) _{RSB díentreÈe (dB)}

Distor sion en sortie (dB) Figur e Hn f n Xn f Hn f Sn f Bn f H1 n f Hn f

3.2. Est imat ion de la propagat ion relat ive de la parole Hn f Least-Mean Square + /-F igur e m∈ M m n fk Hm n fk Hm n − 1 fk µ fk n Xm n∗ fk h Xm n fk − Hm n − 1 fk X1 n fk i µ fk n µ0 SP Pn fk φx1 fk n

SP Pn fk Speech Presence Probability µ0

φx1 fk n

x1 n x1

φx1 fk n βφx1 fk n − − β |X1 n fk |

2

H beamfor ming f φb1 f φb2 f φb f φb1b2 f φbf c f |c f | < bf φbf " c f c f ∗ # H f X1 f H1 f S f B1 f S f B1 f X2 f H2 f S f B2 f H f S f B2 f H f H2 f b H f E X2 f X1 f ∗ E |X1 f |2 H f φs f E B2 f B1 f ∗ φs f φb1 f H f 1 RSB (f ) c f ∗ RSB f H f b H f H f 1 RSB (f ) c f ∗ RSB f RSB f φs f φ_bf H f

3.2. Est imat ion de la propagat ion relat ive de la parole c f GM V D R f b f − 1H fb b H f H bf − 1H fb b H f b f GM V D R − H (f )c(f ) 1+_{R S B ( f )}1 − |c(f )|2 1+ RSB (f ) H (f )− c(f )∗ 1+_{R S B ( f )}1 RSB (f ) 1+ RSB (f ) 2|H f − c f ∗_|2_{− |c f |}2 beamforming D i st or si on D i stoM V D R f E h |S f − GM V D R f HH f S f |2 i φs f E − GM V D R f H " H f # 2 GM V D R f D i stoM V D R f φs f (1− |c(f )|2_{)(1+ R SB (f ))}2 |H (f )− c(f )∗_|2_{RSB (f )} RSB f 2 RSB f |H f − c f ∗ |2 R f |H f − c f ∗ |2 − |c f |2

φs f D i stor f D i stoM V D R f φs f (1+ R SB (f ))2 R(f )R SB (f ) RSB f 2 R(f) (dB) R S B d ’e n tr é e ( d B ) dB −5 0 5 10 15 20 −5 0 5 10 15 20 25 −60 −50 −40 −30 −20 −10 0 F igur e R f R f B r ui t r ési duel B r ui tM V D R f E h |GM V D R f HB f |2 i b H f H b f − 1H fb B r ui tM V D R f φb f RSB f 2|H f − c f ∗|2 RSB f 2 _{− |c f |}2

3.2. Est imat ion de la propagat ion relat ive de la parole R f B r ui tat t f φb f B r ui tM V D R f RSB f 2R f RSB f 2 R f R(f) (dB) R S B d ’e n tr é e ( d B ) −5 0 5 10 15 20 −5 0 5 10 15 20 25 2 4 6 8 10 12 14 16 18 20 dB F igur e R f R SB en sor t i e RSBM V D R f E h |GM V D R f HH f S f |2 i E |GM V D R f HB f |2 RSBM V D R f H fb H b f − 1H fb φs f RSB (f ) 1+ RSB (f )|H − c f ∗ |2− |c f |2 b H f H bf − 1H fb 2

RSBM V D R f − |c f |2 RSB (f )2 (1+ R SB (f ))2 |H (f )− c(f )∗_|2 1− |c(f )|2 RSB (f ) 1+ RSB (f ) > RSB (f )2 (1+ RSB (f ))2 RSB (f ) 1+ RSB (f ) ∈ RSB f RSB f |H f − c f ∗ |2 − |c f |2 ! ₂ > RSB f 2 RSB f 2 |H f − c f ∗ |2 − |c f |2 ! ₂ > RSB f 2 RSB f 2 |H f − c f ∗ |2 − |c f |2 | { z } > 0

L i en avec l ’ acoust i que

R f H f f d H f e − 2iπf Fs d cs R f H f f Fs d cs Broadside d f Fs d cs f Fs d cs d Endfire R f f Fs d cs c f

3.2. Est imat ion de la propagat ion relat ive de la parole C o h È re n c e 1 0 8 0 6 0 4 0 2 0 0 2 0 4 0 6 0 8 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0 1 2 3 4 5 6 7 8 9 10 dB F igur e R f f Fs d cs − |c f |2 R f H f Mean-Squared Coherence R f Broadside Endfire Tabl e R f R f R f R f ∼ R f R f

Broadside 0 1000 2000 3000 4000 5000 6000 7000 8000 −45 −40 −35 −30 −25 −20 −15 −10 −5 Fréquence (Hz) D is to rs io n r e la ti v e ( d B ) 4cm 2cm 0 1000 2000 3000 4000 5000 6000 7000 8000 −1 0 1 2 3 4 5 Fréquence (Hz) G a in e n R S B ( d B ) 4cm 2cm F igur e Wideband

3.2. Est imat ion de la propagat ion relat ive de la parole C oncl usi on Distor tionless R f 10 0 10 20 15 10 5 0 5 0.5 1 1.5 2 2.5 3 Energie relative de la reÈverbeÈration (dB) Gain en RSB (dB) RSB díentreÈe (dB) 5 0 5 10 15 20 15 10 5 0 5 30 25 20 15 10 Energie relative de la

reÈverbeÈration (dB) RSB díentreÈe (dB)

Distor sion en sortie (dB) F igur e

b t fk n Bm n fk √ N ⌊α_{n N}⌋_{+ N} X t =⌊α_{n N}⌋_{+ 1} bm t e− 2iπfkt Bn fk B1 n fk BM n fk T b t N E h Bn fk Bn fk′ H i ∼ _bfk δ_{k k}′ b fk fk n bfk n α bfk n − − α Bn fk Bn fk H α ∈ b fk n b fk n αn fk bfk n − − αn fk Xn fk Xn fk H αn fk α0 − α0 SP Pn fk α0 RSBpost fk n |X1 n(fk)| 2 φ_b1(fk)

3.4. Considérat ions sur le placement des capt eurs φb1 fk RSBpost fk n φb1 fk A l gor i t hm e 1 b fk n i f RSBpost fk n < t hen αn fk α0 − α0 SP P fk n b fk n αn fk bfk n − − αn fk Xn fk XHn fk el se b fk n b fk n − end i f U A B 1 b U A B 2 Tabl e a posteriori beamfor ming Endfire

d F igur e H f αe− 2iπτf Fs τ τ d cs p x2 _{d y} 2_{− d} cs cs α Ren g α Reng Ren p d y 2 _x2 d g s x y+ d 2 !

3.4. Considérat ions sur le placement des capt eurs 0 500 1000 1500 2000 2500 3000 3500 4000 −80 −70 −60 −50 −40 −30 −20 −10 0 Fréquence (dB) D S P ( d B ) Omnidirectionnel Unidirectionnel F igur e Wideband x∈ _{− y}∈ ₋ x y Endfire Broadside −∞

Axe Endfire (cm) A x e B ro a d s id e ( c m ) −1 0 1 2 3 4 −2 0 2 4 −80 −70 −60 −50 −40 dB Axe Endfire (cm) A x e B ro a d s id e ( c m ) −1 0 1 2 3 4 −2 0 2 4 −80 −70 −60 −50 −40 dB 0 0.5 1 1.5 2 2.5 3 3.5 4 −180 −160 −140 −120 −100 −80 −60 −40 −20 Distance (cm) D is to rs io n r e la ti v e ( d B ) Omni − Endfire Omni − Broadside Uni − Endfire Uni − Broadside Endfire Broadside Figur e Broadside

3.4. Considérat ions sur le placement des capt eurs Axe Endfire (cm) A x e B ro a d s id e ( c m ) −1 0 1 2 3 4 −4 −2 0 2 4 −4 −2 0 2 4 dB Axe Endfire (cm) A x e B ro a d s id e ( c m ) −1 0 1 2 3 4 −4 −2 0 2 4 −4 −2 0 2 4 dB 0 0.5 1 1.5 2 2.5 3 3.5 4 −5 −4 −3 −2 −1 0 1 2 3 4 5 Distance (cm) A tt e n u a ti o n d e b ru it ( d B ) Omni − Endfire Omni − Broadside Uni − Endfire Uni − Broadside Endfire Broadside F igur e Endfire

Axe Endfire (cm) A x e B ro a d s id e ( c m ) −1 0 1 2 3 4 −2 0 2 4 −4 −2 0 2 dB Axe Endfire (cm) A x e B ro a d s id e ( c m ) −1 0 1 2 3 4 −2 0 2 4 −4 −2 0 2 dB 0 0.5 1 1.5 2 2.5 3 3.5 4 −4 −3 −2 −1 0 1 2 3 4 5 Distance (cm) G a in e n R S B ( d B ) Omni − Endfire Omni − Broadside Uni − Endfire Uni − Broadside Endfire Broadside F igur e Endfire Broadside

3.5. Synt hèse Endfire Broadside Endfire Broadside Broadside beamfor ming

88 89 91 96 102 104 Adaptive Noise Cancel lation

xm t m∈ t xm t { hsm ⊗s} t bm t hs_m t m s t bm t m hs₁ t δt δ

+

/-F igur e

4.2. Compensat ion de dist orsion xA N C t xA N C t xA N C t { hd⊗s} t | { z } parole fi lt rée bA N C t | { z } bruit résiduel s t xA N C t x1 t xA N C x1 t xA N C t x1 t s t b t υ t xA N C t { hd⊗s} t { hb⊗b} t s t hd t xA N C b t υ t x1 t b t xA N C t { hb⊗b} t xA N C t υ t X1 f S f B f f XA N C f Hd f S f Hb f B f X1 f XA N C f S f B f f x1 t xA N C t s t b t υ t Hd f Hb f hd t

x1 t F igur e S f B f f 3 d Hout f Hd f φs f Hb f φbf |Hd f |2φs f |Hbf |2φbf φs f φb f s t b t S f Hdout f ∗XA N C f S f S f Hdout f ∗ XA N C f Hd f φs f Hbf φbf |Hd f |2φs f |Hb f |2φb f ∗ Hd f S f Hb f B f Ssi g f Hdout f ∗ Hd f S f Hd f φs f Hbf φbf |Hd f |2φs f |Hbf |2φb f ∗ Hd f S f XA N C f RSBA N C f

4.3. Annulat ion de bruit adapt at ive RSBA N C f |Hd f |2φs f |Hbf |2φbf Ssi g f Hb f Hd f RSBA N C f RSBA N C f S f D i sto f E h |S f − Ssi g f |2 i φs f D i sto f − Hb f Hd f RSBA N C f 2 xA N C xA N C x2 t s t b2 t x3 t { hs⊗s} t b3 t b2 t b3 t s t hs t

X2 f S f B2 f X3 f Hs f S f B3 f X2 f X3 f S f B f Hs f x2 t x3 t s t b t hs t X2 f B2 f X3 f B3 f F igur e B2 f B3 f HA N C f φb2b3 f φb2 f X_{A N C}b f B3 f − HA N C f B2 f φb A N C f φb3 f − |HA N C f | 2_φ b2 f φ_b3 f φb2 f b2 b3

4.3. Annulat ion de bruit adapt at ive B3 f b f φb3 f φb A N C f bf φ_b3 f φb3 f − |HA N C f |2φb2 f − |φb2b3 f | 2 φ_b2 f φb3 f |φb2b3 f | 2 φb2 f φb3 f Mean-Squared Coherence B2 f B3 f M SC23 f b f − M SC23 f 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 5 10 15 20 25 30 MSC A tt é n u a ti o n d e b ru it ( d B ) F igur e hA N C

hA N C X_{A N C}si g f Hs f S f − HA N C f ∗ S f S f Hs f − HA N C f φsi g A N C f φs f |Hs f − HA N C f |2 si g f φs f φsi g A N C f |Hs f − HA N C f |2 Hs HA N C HA N C Endfire HA N C Hs |Hs f − HA N C f |2

4.3. Annulat ion de bruit adapt at ive 0 100 200 300 400 500 600 700 800 900 1000 −35 −30 −25 −20 −15 −10 −5 0 5 10 15 Fréquence (Hz) |H s − H a n c | 2 ( d B ) 4cm 8cm 12cm 0 100 200 300 400 500 600 700 800 900 1000 −35 −30 −25 −20 −15 −10 −5 0 5 10 15 Fréquence (Hz) |H s − H a n c | 2 ( d B ) 4cm 8cm 12cm F igur e Endfire RSB f φb2 f φs f φsi g A N C f φb A N C f |Hs f − HA N C f |2 − M SC23 f

MSC |H s − H A N C | 2 (d B ) 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 −30 −25 −20 −15 −10 −5 0 5 0 5 10 15 20 25 dB F igur e |Hs f − HA N C f |2 |Hs f − HA N C f |2 Least-Mean Square

4.4. Implément at ion et st rat égie acoust ique B2 f B f B3 f Hbf B f h f HA N C f Hb f h f XA N C f Hs f − Hb f h f S f − h f B f R SB f |Hs f − Hbf HA N C f |2 | h f |2 6 |Hs f − Hbf | | h f | 2 |Hs f − Hbf | | h f | 2 RSB f | h f |2 |Hs f − Hbf |2 |∆h|2(dB) |H s − H b | 2 (d B ) −40 −35 −30 −25 −20 −15 −10 −40 −35 −30 −25 −20 −15 −10 5 10 15 20 25 30 dB F igur e | h|2 |Hs f − Hb f |2 | h|2

Normalized LMS t x2 t x2 t − L x2 t − L x2 t T L J h E h |x3 t − − hHx2 t |2 i x2 t x3 t ⌊L⌋ ⌊_z⌋ _z z hA N C hA N C t hA N C t µ0 x2 t Hx2 t ǫ x3 t − − hA N C t Hx2 t x2 t µ0 ǫ x2 t Hx2 t ∼ hA N C t hA N C t

4.4. Implément at ion et st rat égie acoust ique 0 5 10 15 −1 −0.5 0 0.5 1 Temps (s) 0 5 10 15 −0.5 0 0.5 1 Temps (s) Signal de parole VAD F igur e L µ0 Endfire

Taille du Filtre ANC P a s d ’a d a p ta ti o n 20 40 60 80 100 120 140 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 9.5 10 10.5 11 11.5 dB

Taille du Filtre ANC

P a s d ’a d a p ta ti o n 20 40 60 80 100 120 140 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 6.5 7 7.5 8 8.5 dB F igur e Endfire

4.4. Implément at ion et st rat égie acoust ique

Taille du Filtre ANC

P a s d ’a d a p ta ti o n 20 40 60 80 100 120 140 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 −10 −8 −6 −4 −2 0 2 4

Taille du Filtre ANC

P a s d ’a d a p ta ti o n 20 40 60 80 100 120 140 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 −10 −8 −6 −4 −2 0 2 4

Taille du Filtre ANC

P a s d ’a d a p ta ti o n 20 40 60 80 100 120 140 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 −10 −8 −6 −4 −2 0 2 4

Taille du Filtre ANC

P a s d ’a d a p ta ti o n 20 40 60 80 100 120 140 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 −10 −8 −6 −4 −2 0 2 4

Taille du Filtre ANC

P a s d ’a d a p ta ti o n 20 40 60 80 100 120 140 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 −10 −8 −6 −4 −2 0 2 4

Taille du Filtre ANC

P a s d ’a d a p ta ti o n 20 40 60 80 100 120 140 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 −10 −8 −6 −4 −2 0 2 4 F igur e Endfire

D i sto P t∈_P s t − ssi g t 2 P t∈_P s t 2 P ssi g t RSB P t∈_P ssi g t 2 P t∈_P sb t 2 , _P t∈_P s t 2 P t∈P b1 t 2 sbt x1

4.5. Performances globales

Taille du Filtre de compensation

P a s d ’a d a p ta ti o n 50 100 150 200 250 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 −9 −8 −7 −6 −5 −4 dB

Taille du Filtre de compensation

P a s d ’a d a p ta ti o n 50 100 150 200 250 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 2.5 3 3.5 4 4.5 5 5.5 dB F igur e

Speech D i st or t i on W ei ght ed

M ult i chan n el W i en er

F i lt er

106 110 115 117 Multichannel Wiener Filter

m M xm t sm t bm t sm t { hm ⊗s} t m w M L w g E h |b1 t − − gHx t |2 i xm t xm t − L xm t T x t x1 t T x2 t T xM t T T sm t bm t s t b t xm L × x t M L × b s1 t − x1 t − − wHx t

5.1. Principe et modèle

+

/-Figur e w E h x t x t T i _{− 1} E x t b1 t − w w g E h |b1 t − − gHb t |2 i | { z } eb E h |gHs t |2 i | { z } es es eb wρ g E h |b1 t − − gHb t |2 i ρE h |gHs t |2 i ρ ρ wρ ρRs t Rb t − 1E b t b ∗ 1 t − Rs t E h s t s t Hi _R b t E h b t b t Hi ρ ρ wρ Rb t − 1Rb t (: L −∆) _{| { z}} (L −∆)-ième posit ion

H

Rbt (: L −∆) L − Rbt

. . . F igur e Rs t E h s t s t H i Rb t E h b t b t H i E b t b∗ 1 t − ρ Voice Activity Detector b t Rbt Rb t ( λbRb t − − λb x t x t H Rb t − λb∈ E b t b∗1 t − L − Rb t < L Rs Rx t E h x t x t H i Rx t Rs t Rb t Rx t Rb t Rx t λxRx t − − λx x t x t H λx ∈ λx ≤ λb Rs t Rx t − Rb t L ρ ρ

5.1. Principe et modèle Jρ w E h |b1 t − − wHb t |2 i ρ E h |wHs t |2 i δJρ w ρRs t Rbt w − E b t b ∗ 1 t − w t w t − − µδJρ w µ x t H_{x t ǫ} µ µ0 x t H_{x t ǫ} ǫ x t Hx t ∼

Minimum Var iance Dis-tor tionless Response

Xn f H f Sn f Bn f Xn f M n f H f Sn f Bn f bf φs f E h |B1 f − W f HB f |2 i ρE h |W f H_{H f S f |}2i W f h ρφs f H f H f H bf i_{− 1} bf (: 1) b f (: 1) bf bf − 1 b f (: 1)

W f − ρφs f ρφs f H f H b f − 1H f | { z } P ost − f i l t r e b f − 1H f H H f H b f − 1H f | { z } M V D R S f X1 f − W f HX f X1 f H X f S f − W f H X f ρφs f ρφs f H f H b f − 1H f | { z } P ost − f i l t r e H f b f − 1 H f H bf − 1H f | { z } M V D R X f ρ beamforming M X1 f H1 f S f B1 f S f B1 f X2 f H2 f S f B2 f H f S f B2 f Xn f H f Sn f Bn f Xn f n f H f Sn f Bn f

5.2. Performances bf b f φb f " c f c f ∗ # φb f φb1 f φb2 f φbf c f c f φb1b2 f φbf φs f H f H f " H f # f G J_ρf G φb1 f − G H " φb1 f φb1b2 f ∗ # − h φb1 f φb1b2 f i G GH ρφs f H f H f H b f G ρ Cb f " φb1 f φb1b2 f ∗ # WM W F f h ρφs f H f H f H bf i− 1 Cbf H f bf Cb f WM W F f " # − ρφs(f ) φb(f )(1− |c(f )|2) " − c f H f H f − c f ∗ # ρφs(f ) φb(f ) |H (f )− c(f )∗_|2 1− |c(f )|2 R f |H f − c f ∗ |2 − |c f |2 R f φs f φbf

D i stoM W F f E h WM W F f HH f S f |2 i φs f E WM W F f H " H f #2 WM W F f D i stoM W F f φs f ρRSB f |H (f )− c(f )_{1− |c(f )|}2∗|2 2 R f |H f − c f ∗ |2 − |c f |2 φs f D i stor f ρRSB f R f 2 R f ρ R(f) (dB) R S B d ’e n tr é e ( d B ) −5 0 5 10 15 20 −5 0 5 10 15 20 25 −90 −80 −70 −60 −50 −40 −30 −20 −10 dB ρ R(f) (dB) R S B d ’e n tr é e ( d B ) −5 0 5 10 15 20 −5 0 5 10 15 20 25 −100 −90 −80 −70 −60 −50 −40 −30 −20 −10 dB ρ F igur e R f ρ ρ ρ R f R f

5.2. Performances B r ui tM W F f E h |B1 f − WM W F f HB f |2 i B r ui tM W F f φb f ρRSB f ρRSB f |H (f )− c(f )_{1− |c(f )|}2∗|2 2 | { z } A |H f − c f ∗|2 − |c f |2 ! | { z } B R f B r ui tat t f φb f B r ui tM W F f ρRSB f R f ρRSB f 2 R f R f ρ R(f) (dB) R S B d ’e n tr é e ( d B ) −5 0 5 10 15 20 −5 0 5 10 15 20 25 5 10 15 20 25 dB ρ R(f) (dB) R S B d ’e n tr é e ( d B ) −5 0 5 10 15 20 −5 0 5 10 15 20 25 5 10 15 20 25 dB ρ F igur e R f ρ ρ R f R f RSBM W F f E h |S f − WM W F f HH f S f |2 i E |B1 f − WM W F f HB f |2

M W F − |c f |2 ρ R f RSB f RSBM W F f RSB f R f R(f) (dB) R S B d ’e n tr é e ( d B ) −5 0 5 10 15 20 −5 0 5 10 15 20 25 5 10 15 20 25 dB F igur e R f R f R f R f

5.3. Considérat ions sur le placement des capt eurs ρ x ∈ _{− y} ∈ ₋ x y Endfire Broadside −∞ Axe Endfire (cm) A x e B ro a d s id e ( c m ) 0 5 10 15 20 −20 −10 0 10 20 −8.5 −8 −7.5 −7 −6.5 −6 dB 0 2 4 6 8 10 12 14 16 18 20 −9 −8.5 −8 −7.5 −7 −6.5 −6 −5.5 Distance (cm) D is to rs io n r e la ti v e ( d B ) Endfire Broadside F igur e ρ Endfire

Axe Endfire (cm) A x e B ro a d s id e ( c m ) 0 5 10 15 20 −10 0 10 20 4 5 6 7 dB 0 2 4 6 8 10 12 14 16 18 20 3 3.5 4 4.5 5 5.5 6 6.5 7 Distance (cm) A tt é n u a ti o n d u b ru it ( d B ) Endfire Broadside F igur e ρ Endfire Axe Endfire (cm) A x e B ro a d s id e ( c m ) 0 5 10 15 20 −20 −10 0 10 20 1 2 3 4 5 6 dB 0 2 4 6 8 10 12 14 16 18 20 −1 0 1 2 3 4 5 6 7 Distance (cm) G a in e n R S B ( d B ) Endifre Broadside F igur e ρ Endfire Endfire Endfire

5.4. Synt hèse

Minimum Variance Distortionless Response 121

125

M V D R adapt at i f - C hapi t r e 3 C R -A N C - C hapi t r e 4 SD W -M W F - C hapi t r e 5 Wideband Broadside Wideband Tr a i t e m e n t m o n o c a n a l Tr a i t e m e n t h y b r i d e Ba n c d ' a n a l y se B a n c d e sy n t h Ëse En t r Èe s So r t i e

+

F igur e Wideband

6.1. Syst ème MVDR + CR-ANC

4cm

4cm

4cm

o m n i d i r e c t i o n n e l s u n i d i r e c t i o n n e l s Lo cu t e u r F igur e Broad-side Endfire un banc de fi l t r e un t r ai t em ent différ enci é

Crosstalk Resistant Adaptive Noise Cancel lation

une r econst r uct i on fr équent i el l e

F igur e

Mean-Squared Coherence

Adaptive Noise Cancel lation

6.1. Syst ème MVDR + CR-ANC Broadside Endfire

CR-A N C

M V D R

F igur e B anc de fi l t r es hbf t hhf t L hbf t hhf t δ t − L L

0 500 1000 1500 2000 2500 3000 3500 4000 −200 −180 −160 −140 −120 −100 −80 −60 −40 Fréquence (Hz) R é p o n s e e n f ré q u e n c e ( _Passe−haut Passe−bas 0 2 4 6 8 10 12 14 16 −0.2 0 0.2 0.4 0.6 0.8 Temps (ms) R é p o n s e i m p u ls io n n e ll e F igur e C R -A N C Nor malized LMS µ0 µ0

Voice Activity Detector

a posteriori A l gor i t hm e 2 i f N P _{N − 1} k= 0 RSBpost fk n < C R− A N C t hen V AD el se V AD end i f C R− A N C C R− A N C M V D R

6.2. Syst ème MVDR + SDW-MWF

µ0

β

α0

Speech Presence Probability a posteriori

M V D R SD W -M W F H y br i de Perceptual Evaluation of Speech Quality Narrowband seg M M − 1_X m = 0 10 P _{N m + N − 1} t = N m s t 2 P _{N m + N − 1} t = N m s t − s t 2 M N s t s t

Mean Opinion Score

6.2. Syst ème MVDR + SDW-MWF −5 0 5 10 1.6 1.8 2 2.2 2.4 2.6 2.8 3 RSB en entrée (dB) P E S Q e n s o rt ie ( é c h e ll e M O S ) Hybride MVDR SDW−MWF F igur e −50 0 5 10 1 2 3 4 5 6 7 RSB en entrée (dB) G a in e n R S B s e g m e n ta l (d B ) Hybride MVDR SDW−MWF F igur e

A nt enne acoust i que

Broadside