CS5240

download CS5240

of 6

Transcript of CS5240

  • 8/12/2019 CS5240

    1/6

    Gender Classification from Speech

    Chiu Ying Lay

    Ng Hian James

    Abstract

    his p!ojec" uses #$L$% "o de&ise a

    gende! c'assi(ie! (!om speech )y ana'y*ing

    "he &oice samp'es con"aining an a!)i"!a!y

    sen"ence. he speech signa' is assumed "ocon"ain on'y + spea,e!- spea,ing in ng'ish-

    /i"h no o"he! )ac,g!ound sounds. he

    c'assi(ie! ana'yses "he &oice samp'es )y

    using a pi"ch de"ec"ion a'go!i"hm )ased oncompu"ing "he sho!""ime au"oco!!e'a"ion

    (unc"ion o( "he speech signa'.

    1. Introduction

    he u'"ima"e goa' in au"oma"ic speech

    !ecogni"ion is "o p!oduce a sys"em /hich

    can !ecogni*e con"inuous speech u""e!ances

    (!om any spea,e! o( a gi&en 'anguage. 1ne

    o( "he main app'ica"ion a!eas (o! speech

    !ecogni"ion is &oice inpu" "o compu"e!s (o!such "as,s as documen" c!ea"ion /o!dp!ocessing and (inancia' "!ansac"ion

    p!ocessing "e'ephone)an,ing. $u"oma"ic

    speech !ecogni"ion is done in pa!"s /i"h

    gende! c'assi(ica"ion. he need (o! gende!c'assi(ica"ion (!om speech a'so a!ises in

    se&e!a' si"ua"ions such as so!"ing "e'ephone

    ca''s )y gende! eg. gende! sensi"i&e

    su!&eys- as pa!" o( an au"oma"ic speech

    !ecogni"ion sys"em "o enhance spea,e!

    adap"a"ion and as pa!" o( au"oma"ic spea,e!

    !ecogni"ion sys"ems.

    peech sounds can )e di&ided in"o "h!ee

    )!oad c'asses acco!ding "o "he mode o(

    e5ci"a"ion. he "h!ee c'asses a!e &oiced

    sounds- un&oiced sounds and p'osi&e

    sounds. $" a 'inguis"ic 'e&e'- speech can )e

    &ie/ed as a se6uence o( )asic sound uni"sca''ed phonemes. he same phoneme may

    gi&e !ise "o many di((e!en" sounds o!

    allophonesa" "he acous"ic 'e&e'- dependingon "he phonemes /hich su!!ound i".

    7i((e!en" spea,e!s p!oducing "he same s"!ing

    o( phonemes con&ey "he same in(o!ma"ion

    ye" sound di((e!en" as a !esu'" o( di((e!ences

    in dia'ec" and &oca' "!ac" 'eng"h and shape.

    Li,e mos" 'anguages- ng'ish can )e

    desc!i)ed in "e!ms o( a se" o( 40 o! so

    phonemes o! a!"icu'a"o!y ges"u!es +9.

    Nea!'y a'' in(o!ma"ion in speech is in "he

    !ange 200H* "o 8,H*. Humans disc!imina"e

    &oices )e"/een ma'es and (ema'es acco!ding

    "o "he (!e6uency. :ema'es spea, /i"h highe!

    (undamen"a' (!e6uencies "han ma'es. he

    adu'" ma'e is (!om a)ou" ;0H* "o 2;0H*-

    /i"h an a&e!age &a'ue o( a)ou" +20H*. :o!

    an adu'" (ema'e- "he uppe! 'imi" o( "he !ange

    is o( much highe!- pe!haps as high as ;00H*.he!e(o!e- )y ana'y*ing "he a&e!age pi"ch o(

    "he speech samp'es- /e can de!i&e ana'go!i"hm (o! a gende! c'assi(ie!.

    o p!ocess a &oice signa'- "he!e a!e

    "echni6ues "ha" can )e )!oad'y c'assi(ied as

    ei"he! "imedomain o! (!e6uencydomain

    app!oaches. i"h a "imedomain app!oach-

    in(o!ma"ion is e5"!ac"ed )y pe!(o!ming

    measu!emen"s di!ec"'y on "he speech signa'

    /he!eas /i"h a (!e6uencydomain app!oach-

    "he (!e6uency con"en" o( "he signa' is

    ini"ia''y compu"ed and in(o!ma"ion is

    e5"!ac"ed (!om "he spec"!um. =i&en suchin(o!ma"ion- /e can pe!(o!m ana'ysis on "he

    di((e!ences in pi"ch- *e!oc!ossing !a"e

    C? and (o!man" posi"ions (o! &o/e's

    )e"/een ma'e and (ema'e.

    his pape! is o!gani*ed as (o''o/s sec"ion 2

    gi&es a 'is" o( di((e!en" (ea"u!e e5"!ac"ion

    me"hods as /e'' as c'assi(ica"ion "echni6ues

  • 8/12/2019 CS5240

    2/6

    /hi'e sec"ion 3 is a)ou" ou! imp'emen"a"ion

    o( a gende! c'assi(ie!. ec"ion 4 p!esen"s ou!

    e&a'ua"ion o( "he imp'emen"ed c'assi(ie! and

    sec"ion ; "ouches on some p!oposed idea (o!

    (u"u!e enhancemen"s.

    2. Classification Techniques

    he di((e!en" (ea"u!es o( a speech "ha" can )e

    e5"!ac"ed (o! ana'ysis a!e )asica''y (o!man"

    (!e6uency and pi"ch (!e6uency. %ased on ou!

    su!&ey in"o "he cu!!en" 'i"e!a"u!e- &a!ious

    imp'emen"a"ions ha&e )een done using "he

    a)o&emen"ioned (ea"u!es "o c'assi(y &oice

    samp'es acco!ding "o gende!. he (o''o/ing

    su)sec"ions high'igh" "he &a!ious "echni6ues

    o( speech (ea"u!e e5"!ac"ion.

    2.1.Pitch Analysis

    Ai"ch is de(ined as "he (undamen"a'

    (!e6uency o( "he e5ci"a"ion sou!ce. Hence an

    e((icien" pi"ch e5"!ac"o! and an accu!a"e

    pi"ch es"ima"e ca'cu'a"ed can )e used in an

    a'go!i"hm (o! gende! iden"i(ica"ion. hepape!s /e su!&eyed p!o&ide mu'"ip'e aspec"s

    in e5"!ac"ing and es"ima"ing pi"ch (o! gende!

    c'assi(ica"ion.

    =o'd?a)ine! a'go!i"hm 29 i''us"!a"es pi"ch

    e5"!ac"ion )ased on "he (ac" "ha" 'oca"ing "he

    posi"ion o( "he ma5imum poin" o( e5ci"a"ion

    is no" a'/ays de"e!mina)'e (!om "he "ime

    /a&e(o!m. he!e(o!e i" uses addi"iona'

    (ea"u!es o( "he "ime/a&e(o!m "o o)"ain a

    num)e! o( pa!a''e' es"ima"es o( "he pi"ch

    pe!iod- as /e'' as de"ec"ing "he pea, signa'

    &a'ues.

    e&e!a' /o!,s ha&e imp'emen"ed pi"ch

    e5"!ac"ion a'go!i"hms )ased on compu"ing

    "he sho!""ime au"oco!!e'a"ion (unc"ion o(

    "he speech signa'. :i!s"- "he speech isno!ma''y 'o/passed (i'"e!ed a" a (!e6uency

    o( a)ou" +,H*- /hich is /e'' a)o&e "he

    ma5imum an"icipa"ed (!e6uency !ange (o!

    pi"ch. :i'"e!ing he'ps "o !educe "he e((ec"s o(

    "he highe! (o!ma"s and any e5"!aneous high

    (!e6uency noise. he signa' is /indo/ed

    using an app!op!ia"e so(" /indo/ such as

    Hamming o( du!a"ion 20 "o 30 ms and a

    "ypica' au"oco!!e'a"ion (unc"ion is gi&en )y

    =

    +=

    n

    knxnxkR 989.832

    he au"oco!!e'a"ion (unc"ion gi&es ameasu!e o( "he co!!e'a"ion o( a signa' /i"h a

    de'ayed copy o( i"se'(. Bn "he case o( &oiced

    speech- "he main pea, in sho!""ime

    au"oco!!e'a"ion (unc"ion no!ma''y occu!s a"a 'ag e6ua' "o "he pi"chpe!iod. his pea, is

    "he!e(o!e de"ec"ed and i"s "ime posi"ion gi&es

    "he pi"ch pe!iod o( "he inpu" speech.

    $("e! e5"!ac"ing pi"ch in(o!ma"ion (!om

    speech (i'es- pi"ch es"ima"ion a'go!i"hm is

    "hen usua''y app'ied. $ &e!sion o( "he pi"ch

    es"ima"ion a'go!i"hm used (o! B#% speech

    coding as desc!i)ed in 39 gi&es an a&e!age

    pi"ch es"ima"e (o! "he spea,e! )y es"ima"ing

    "he pi"ch (o! each (!ame o( "he speech. $n

    ini"ia' es"ima"e o( "he a&e!age pi"ch /as

    ca'cu'a"ed ac!oss "he !egions o( in"e!es"

    iden"i(ied )y a pa""e!n ma"che!. he es"ima"eis !e(ined )y ca'cu'a"ing a ne/ a&e!age (!om

    pi"ch es"ima"es /i"hin a pe!cen"age o( "he

    o!igina' a&e!age. hus "his !emo&es "he

    ou"'ie!s p!oduced )y pi"ch dou)'ing- "!ip'ing

    and e!!o! in !egion c'assi(ica"ion. his

    "echni6ue using pi"ch can )e used in

    iso'a"ion (o! gende! iden"i(ica"ion )ycompa!ing "he a&e!age pi"ch es"ima"e /i"h

    p!ese" "h!esho'd. s"ima"es )e'o/ "he

    "h!esho'd a!e iden"i(ied as ma'e and "hose

    a)o&e as (ema'e.

    $n a'"e!na"i&e "echni6ue in pi"ch ana'ysis is

    )y 'oo,ing a" "he *e!oc!ossing !a"e C?

    and sho!""ime ene!gy (unc"ion o( a speech

    (i'e 49. C? is a measu!e o( "he num)e! o(

    "imes in a gi&en "ime in"e!&a' (!ame "ha"

    "he amp'i"ude o( "he speech signa' passes

    "h!ough "he *e!oa5is. C? is an impo!"an"pa!ame"e! (o! &oicedun&oiced c'assi(ica"ion

    and endpoin" de"ec"ion as /e'' as gende!

    c'assi(ica"ion as "he C? (o! (ema'e &oice is

    highe! "han "ha" (o! ma'e &oice. he sho!"

    "ime ene!gy (unc"ion o( speech is compu"ed

    )y sp'i""ing "he speech signa' in"o (!ames o(

    N samp'es and compu"ing "he "o"a' s6ua!ed

  • 8/12/2019 CS5240

    3/6

    &a'ues o( "he signa' samp'es in each (!ame.

    p'i""ing "he signa' in"o (!ames can )e

    achie&ed )y mu'"ip'ying "he signa' )y a

    sui"a)'e /indo/ n9- nD0- +- 2E- N+-

    /hich is *e!o (o! n ou"side "he !ange 0- N+. $ simp'e (unc"ion gi&en "o e5"!ac" a

    measu!e !e'a"ed "o ene!gy can )e de(ined as

    =n

    mnWnxnW 989898

    he ene!gy o( "he &oiced speech is gene!a''y

    g!ea"e! "han "ha" o( un&oiced speech.

    =i&en in 49- "he p!oposed &a!ia)'e "o do

    gende! c'assi(ica"ion is de(ined )y a (unc"ion

    comp!ising "he mean o( C? and "he cen"e!

    o( g!a&i"y o( "he acous"ic &ec"o!. he 'ogic is"ha" "he cen"e! o( g!a&i"y (o! a ma'e &oice

    spec"!um is c'ose! "o 'o/ (!e6uencies and

    "ha" o( (ema'e is "o highe! (!e6uencies.

    =

    =

    =

    f

    fX

    ZCRMeanX

    X

    W f

    f

    f

    f

    f

    f

    2

    +40

    3;

    ;

    +

    /he!e #eanC? is "he mean o( C? in +s

    and F( is (!e6uency coe((icien" o( G(. he

    shou'd )e highe! (o! ma'e &oices.

    2.2. Formant Analysis

    $ (o!man" is a dis"inguishing o! meaning(u'

    (!e6uency componen" o( human speech. B" is

    "he cha!ac"e!is"ic ha!monic "ha" iden"i(ies

    &o/e's "o "he 'is"ene!. his (o''o/s (!om "he

    de(ini"ion "ha" "he in(o!ma"ion humans

    !e6ui!e "o dis"inguish )e"/een &o/e's can )e

    !ep!esen"ed pu!e'y 6uan"i"a"i&e'y )y "he

    (!e6uency con"en" o( "he &o/e' sounds.

    he!e(o!e- (o!man" (!e6uencies a!e

    e5"!eme'y impo!"an" (ea"u!es and (o!man"

    e5"!ac"ion is "hus an impo!"an" aspec" o(speech p!ocessing.

    ince ma'e and (ema'e ha&e di((e!en"

    (o!man" posi"ions (o! &o/e's- "he!e(o!e(o!man" posi"ions can )e used "o de"e!mine

    "he gende! o( a spea,e!. hus "he dis"inc"ion

    )e"/een ma'e and (ema'e cou'd )e

    !ep!esen"ed )y "he 'oca"ion in "he (!e6uency

    domain o( "he (i!s" 3 (o!man"s (o! &o/e's.

    Ie!gin e" a'. ;9 p!esen"ed "ha" an au"oma"ed

    ma'e(ema'e c'assi(ica"ion can )e )ased on

    jus" "he di((e!ence o( "he (i!s" and second

    (o!man"s )e"/een ma'e and (ema'e &oicesamp'es. $ !o)us" )u" (as" a'go!i"hm can

    "hen )e de&e'oped "o de"ec" "he gende! o( aspea,e!.

    hen "a',ing a)ou" using (o!man" ana'ysis

    (o! doing gende! c'assi(ica"ion- "he p!o)'em

    is )asica''y )!o,en do/n "o "/o pa!"s. he

    (i!s" pa!" is (o!man" e5"!ac"ion /hich ;9

    uses a "echni6ue "ha" pe!(o!ms a de"ec"ion o(

    ene!gy concen"!a"ion ins"ead o( "he c'assic

    pea, pic,ing "echni6ue. he second pa!" is

    "he ma'e(ema'e de"ec"ion )ased on "he

    'oca"ion o( "he (i!s" and second (o!man".

    he!e a!e &a!ious /ays p!oposed in "he

    'i"e!a"u!e o( speech p!ocessing (o! e5"!ac"ing

    (o!man"s- especia''y (o! "he (i!s" "/o

    (o!man"s. hough &o/e's can )edis"inguished )y "he (i!s" "h!ee (o!man"s- "he

    "hi!d does no" p'ay an impo!"an" !o'e as i"

    does no" inc!ease "he pe!(o!mance o( any

    c'assi(ie! signi(ican"'y. cha(e! and ?a)ine!

    9 ga&e a pea, pic,ing "echni6ue "ha" has

    )ecome a c'assic )u" 'a"e! s"udies e&a'ua"ed

    i" "o )e s'o/ and inaccu!a"e "o a ce!"ain

    e5"en". hey su)se6uen"'y do ha&e enhanceda'go!i"hms K9 )u" /e did no" s"udy in"o

    "hem and hence no" desc!i)e "hem he!e.

    he mode!n (o!ms o( (o!man" e5"!ac"ion

    s"udied ma,e use o( "he concen"!a"ion o(

    spec"!a' ene!gy "o "!ac, and es"ima"e "he (i!s""/o (o!man"s- as sho/n )y Ie!gin e" a'. and

    Chan/oo im e" a'. 89. Ie!gin e" a'. (i!s"

    de(ine a spec"!a' ene!gy &ec"o! o)"ained

    (!om (as" :ou!ie! "!ans(o!m. hen "oes"ima"e "he (i!s" (o!man"- an ini"ia' in"e!&a'

    )e"/een "/o (!e6uency posi"ions &a'id (o!ma'e and (ema'e is (i5ed. he in"e!&a'

    chosen is )e"/een +2;H* and 8K;H*. he

    'o/e! )ound is inc!eased o! "he uppe! )ound

    is dec!eased )y a (i5ed amoun" in an

    a'go!i"hm un"i' "he di((e!ence !eaches a

    p!ede(ined &a'ue. :ina''y- "he mean posi"ion

    o( "he ene!gy in "he in"e!&a' is es"ima"ed "o

    ge" "he (i!s" (o!man". he second (o!man" is

  • 8/12/2019 CS5240

    4/6

    simi'a!'y (ound /i"h a di((e!en" ini"ia'

    in"e!&a' "ha" is )e"/een "he ma5imum (i!s"

    (o!man" M 2;0H*- 8K;H* and 28K;H*.

    $ 'is" o( "he a&e!age (o!man" (!e6uencies (o!ng'ish &o/e's )y ma'e and (ema'e spea,e!s

    has )een o)"ained )e(o!ehand. :o! a &oicesamp'e- "/o sco!es- co!!esponding "o "he

    num)e! o( "imes "he (o!man" posi"ions o( a

    (!ame a!e assigned ma'e and (ema'e &a'ues.

    o do "his- "he (o!man" 'oca"ions o( "he

    &o/e' (!ames a!e compa!ed /i"h "he

    !e(e!ence ma'e(ema'e (o!man" 'oca"ions o(

    a'' &o/e's. he 'eas" di((e!ence p!o&ides "he

    gende! associa"ed "o "his (!ame. he

    co!!esponding sco!e is inc!eased )y +. $" "he

    end o( "he compu"a"ion- "he g!ea"e! sco!e

    de"e!mines "he es"ima"ed gende! o( "he

    &oice.

    . Implementation

    he mode' "ha" /e ha&e chosen (o!

    imp'emen"a"ion is using pi"ch e5"!ac"ion &ia

    au"oco!!e'a"ion since human ea!s main'ydi((e!en"ia"e )y pi"ch. e ha&e assumed a

    =aussian dis"!i)u"ion "o compu"e "he one

    "ai'ed con(idence in"e!&a' a" O "o assign

    /eigh"s "o "he !esu'"s. %y using one"ai'ed

    con(idence in"e!&a'- /e a'so imp'ied "ha"on'y human speech samp'es /i"hou"

    )ac,g!ound noise a!e supp'ied (o! "!aining

    and gende! de"ec"ion.

    he mode' is imp'emen"ed using #$L$%.

    he!e a!e )asica''y "/o modu'es- Ai"ch

    pi"ch.m and Ai"ch $u"oco!!e'a"ion

    pi"chaco!!.m (o! pi"ch e5"!ac"ion and

    es"ima"ion 9.

    he a'go!i"hm in Ai"ch pi"ch.m (o! pi"ch

    e5"!ac"ion is as (o''o/s

    + he speech is di&ided in"o 0ms

    (!ame segmen"s. ach segmen" is

    e5"!ac"ed a" e&e!y ;0ms in"e!&a'.

    his imp'ies "ha" "he o&e!'ap

    )e"/een segmen"s is +0ms.

    2 ach segmen" ca''s Ai"ch

    $u"oco!!e'a"ion "o es"ima"e "he

    (undamen"a' (!e6uency (o! "ha"

    segmen".

    3 #edian (i'"e!ing is done (o! e&e!y 3

    segmen"s so "ha" i" is 'ess a((ec"ed

    )y noise.4 :ina''y "he a&e!age o( a''

    (undamen"a' (!e6uencies is !e"u!ned.

    he pi"ch es"ima"ed (o! each 0ms (!ame

    segmen" can )e p!esen"ed in a pi"ch con"ou!

    diag!am. B" i''us"!a"es "he pi"ch &a!ia"ion (o!

    "he /ho'e in"e!&a' o( ;s- as sho/n in :igu!e

    +.

    Fi!ure 1" Pitch contour for F1.#a$

    he a'go!i"hm in Ai"ch $u"oco!!e'a"ion

    pi"chaco!!.m (o! pi"ch es"ima"ion using

    au"oco!!e'a"ion "echni6ue is as (o''o/s

    + he speech is no!ma''y 'o/pass

    (i'"e!ed using a 4"ho!de! %u""e!/o!"h

    'o/pass (i'"e! a" (!e6uency o(

    00H* /hich is /e'' a)o&e "he

    ma5imum an"icipa"ed (!e6uency (o!

    pi"ch. he %u""e!/o!"h (i'"e! is a!easona)'e choice "o use as i" is

    app!o5ima"es an idea' 'o/ pass (i'"e!

    as "he o!de! inc!eases.

    2 7ue "o "he compu"a"iona' in"ensi"yo( "he many mu'"ip'ica"ions !e6ui!ed

    (o! "he compu"a"ion o( "he

    au"oco!!e'a"ion (unc"ion- cen"!e

    c'ipping "echni6ue is app'ied "o

    e'imina"e "he need (o! mu'"ip'ica"ion

    in au"oco!!e'a"ion)ased a'go!i"hm.

    his in&o'&es supp!essing &a'ues o(

    "he signa' )e"/een "/o adjus"a)'e

    c'ipping "h!esho'ds. B" is se" a" 0.8

  • 8/12/2019 CS5240

    5/6

    o( "he ma5imum amp'i"ude &a'ue.

    Cen"!ec'ipping !emo&es mos" o( "he

    (o!man" in(o!ma"ion- 'ea&ing

    su)s"an"ia' componen"s due "o "he

    pi"ch pe!iodici"y /hich sho/s upmo!e c'ea!'y in "he au"oco!!e'a"ion

    (unc"ion.3 $("e! c'ipping- "he sho!""ime ene!gy

    (unc"ion is compu"ed. e de(ine

    si'ence i( ma5imum au"oco!!e'a"ion

    is 'ess "han 40O o( "he sho!""ime

    ene!gy. he ma5imum

    au"oco!!e'a"ion is "a,en (!om "he

    !ange o( 0H* "o 320H*. Hence i(

    (undamen"a' (!e6uency (ound

    ou"side "he !ange- i" is "!ea"ed as

    un&oiced segmen".

    %. Trainin!

    igh" pai!s o( &oice samp'es a pai! consis"so( a ma'e and a (ema'e a!e co''ec"ed (o! "he

    "!aining o( "he gende! speech c'assi(ie!. $

    &oice samp'e is assumed "o con"ain on'y +

    spea,e! spea,ing an a!)i"!a!y ng'ishsen"ence (o! ;s /i"hou" )ac,g!ound sounds.

    $cco!ding Ny6uis"Ps samp'ing "heo!em- i(

    "he highes" (!e6uency componen" p!esen" in

    "he signa' is fh H*- "hen "he samp'ing

    (!e6uencyfsmus" )e a" 'eas" "/ice "his &a'ue-"ha" is fs 2fh- in o!de! "o a&oid a'iasing.

    ach samp'e is !eco!ded a" 22.0; ,H* /hich

    is /e'' a)o&e "he "/ice o( 8 ,H* "he highes"

    (!e6uency o)se!&ed (o! speech.

    he a&e!age (undamen"a' (!e6uencies pi"ch

    a!e compu"ed (o! )o"h ma'e c'ass and (ema'e

    c'ass. $ "h!esho'd is o)"ained )y ge""ing "he

    mean o( "he 2 a&e!age (undamen"a'

    (!e6uencies. he s"anda!d de&ia"ion 7 (o!

    each c'ass is a'so compu"ed. he &a'ues a!e

    used as pa!ame"e!s o( "he c'assi(ie! as sho/n)e'o/.

    #ean pi"ch (o! ma'e +4.;+44 H*

    7 (o! ma'e 23.838 H*

    #ean pi"ch (o! (ema'e 2+2.3+34 H*

    7 (o! (ema'e +K.0;3+ H*

    h!esho'd +K.4+3 H*

    he "h!esho'd is "he de"e!minan" (o! "he

    gende! c'ass. B( "he pi"ch o( a &oice samp'e

    (a''s )e'o/ "he "h!esho'd- "he c'assi(ie! /i''

    assign i" as ma'e. 1"he!/ise- i" /i'' assign as(ema'e.

    $ one"ai'ed O con(idence 'e&e' is

    compu"ed "o !e('ec" "he p!o)a)i'i"y o(

    misc'assi(ica"ion. B( i" (a''s ou"side

    con(idence in"e!&a' i.e. i" )e'ongs "o "he

    noncon(iden" !egion- i" is !ema!,ed as

    G#isc'assi(ica"ion possi)'e.

    &. 'esults

    i5 mo!e &oice samp'es a!e "a,en (o! "es"ing

    o( "he gende! speech c'assi(ie!. :i&e o( "hem2 ma'es and 3 (ema'es a!e c'assi(ied

    co!!ec"'y in"o gende! c'asses. Ho/e&e!- one

    o( "he co!!ec"'y c'assi(ied samp'es (a''sou"side "he O con(idence 'e&e'.

    1ne ma'e &oice samp'e is misc'assi(ied in"o

    (ema'e c'ass due "o "he p!esence o( high(!e6uency noise componen". he noise

    componen" gi&es !ise "o a highe!

    (undamen"a' (!e6uency pi"ch- hence i" (a''s

    in"o "he /!ong gende! c'ass /i"h high

    con(idence. he!e(o!e i" is c!i"ica' "o !eco!d&oice samp'e /i"hou" )ac,g!ound o! s"a"ic

    noise.

    (. Future )nhancements

    :!om ou! !esu'"s gi&en in "he a)o&e sec"ion-

    ou! c'assi(ie! )ased on pi"ch e5"!ac"ion using

    au"oco!!e'a"ion managed "o pe!(o!m

    sa"is(ac"o!i'y. Ho/e&e!- "he!e a!e &oice

    samp'es "ha" (ai'ed "o (a'' /i"hin "he !ange o(

    con(idence 'e&e'. Hence "hey canno" )ec'assi(ied /i"h ce!"ain"y. 5"!eme cases o(

    ma'es &oice /i"h highe! pi"ch o! (ema'e

    &oices /i"h 'o/e! pi"ch a!e c'assi(ied in"o

    "he /!ong gende!. his "ype o( si"ua"ions

    can ha!d'y )e imp!o&ed as "he "h!esho'd /e

    de!i&ed has )een c!ossed. e may (ine"une

  • 8/12/2019 CS5240

    6/6

    "he "h!esho'd )y "!aining /i"h a )igge!

    samp'e se".

    1"he! cases o( inaccu!a"e !esu'"s in&o'&e

    &oice samp'es "ha" a!e )eing co!!ec"'yc'assi(ied )u" (a'' in "he noncon(iden"

    !egion. Bmp!o&emen"s can )e made "o hand'esuch cases )y using Gcom)oc'assi(ie!. $

    Gcom)oc'assi(ie! is a c'assi(ie! consis"s o(

    mu'"ip'e c'assi(ie!s emp'oying di((e!en"

    me"hods o( doing gende! de"ec"ion. $ simp'e

    /eigh"sco!ing a'go!i"hm de"e!mines "he

    gende! o( a &oice samp'e )y 'oo,ing a" "he

    !esu'"s !e"u!ned (!om "he g!oup o(

    c'assi(ie!s.

    B" /o!,s in "he (o''o/ing /ay

    + ach c'assi(ie! assigns /eigh" "o "he

    !esu'" )ased on ho/ con(iden" i" iso( "he !esu'"s. :o! e5amp'e- ou!

    imp'emen"a"ion /i'' assign &a!ying

    /eigh"s acco!ding "o "he dis"ance

    a/ay (!om "he mean. B( "he !esu'"(a''s ou"side "he con(idence 'e&e'- a

    (u!"he! discoun"ed /eigh" may )e

    gi&en ins"ead.

    2 he /eigh"s (!om "he c'assi(ie!s a!e

    summed up and "he gende! c'ass "ha"

    has "he highes" sco!e is "a,en as "he

    c'ass. $n a!)i"!a!y "h!esho'd (o! "he

    "o"a' /eigh" can a'so )e de(ined so"ha" "he!e is s"i'' a g!ey a!ea /he!e

    "he c'assi(ica"ion is deemed non

    con(iden".

    *. Conclusions

    Bn "his p!ojec"- /e ha&e imp'emen"ed a

    gende! speech c'assi(ie! )ased on pi"chana'ysis. o sho/ "he su!eness o( ou!

    !esu'"s- a O con(idence 'e&e' is used "o

    demons"!a"e ho/ con(iden" "he c'assi(ie! iso( "he !esu'"s. %ased on ou! !esu'"s- /e canconc'uded "ha" pi"ch di((e!en"ia"ion is an

    e5ce''en" /ay o( c'assi(ying speech in"o "he

    gende! c'asses.

    e a'so p!oposed a Gcom)oc'assi(ie! "ha"

    uses o"he! "echni6ues such as (o!man"

    ana'ysis "o imp'emen" a /eigh"sco!ing

    sys"em so "ha" "he gende! speech

    c'assi(ica"ion is mo!e !o)us". Con(idence

    'e&e' compu"a"ion can )e used (o!

    assignmen" o( /eigh"s.

    'eferences

    +9 :. J. 1/ens- Signal Processing of

    Speech.

    29 =o'd- %. and ?a)ine!- L.?- Parallel

    processing techniques for estimating

    pitch periods of speech in time-domain.

    39 .. Aa!!is and #.J. Ca!ey- Language

    ndependent !ender dentification.

    49 H. Ha!)- L. Chen- J. $u'oge- Speech"Music" Silence and !ender #etection

    $lgorithm.

    ;9 ?. Ie!gin- $. :a!ha"- 7. 1Phaughnessy-

    Ro%ust !ender-#ependent $coustic-

    Phonetic Modelling in Continuous

    Speech Recognition &ased 'n $ (e)$utomatic Male"*emale Classification.

    9 ?.. cha(e! and L.?. ?a)ine!- S+stem

    for automatic formant anal+sis of ,oiced

    speech.

    K9 L.?. ?a)ine! and ?.. cha(e!- #igital

    Processing of Speech Signals.

    89 Chan/oo im and onyong ong-

    o)el Pronunciation $ccurac+

    Checking S+stem &ased on Phoneme

    Segmentation and *ormants .xtraction.

    9 .7. ''is- #esign of a Speaker

    Recognition Code using M$/L$&.