Comparative study of various techniques to measure speech intelligibility

COMPARATIVE STUDY OF VARIOUS TECHNIQUES TO MEASURE SPEECH INTELLIGIBILITY PACS REFERENCE: 43.55.Hy Galindo, Miguel; Zama eño, Teó ilo; Gi ón, Sa a Dep. o Applied Physics II; High School o A chi ec u e; Se ille Uni e si y A da. Reina Me cedes, 2 41012 - Se ille Spain Tel.: +34 954556612 Fax: +34 954557892 E-mail: [email p o ec ed] ABSTRACT The aim o his pape is o compa e wo di e en me hods o measu ing he Modula ion T ans e Func ions (MTF) and he co esponding de i ed STI-RASTI indices o assessing in elligibili y. The i s me hod uses, as es signal, a band il e ed (500 and 2000 Hz) ampli ude modula ed noise. The second one uses MLS signals o ob ain he impulse esponse o he sys em; u he pos p ocessing allows us o calcula e he MTF and STI-RASTI indices om his il e ed esponse. INTRODUCTION The Speech T ansmission Index (STI) and he Rapid Speech T ansmission Index (RASTI) we e p oposed by Hou gas and S eeneken [1]-[3] o e alua e speech in elligibili y, whe e RASTI is a educed e sion o STI. Bo h a e calcula ed om he Modula ion T ans e Func ion (MTF) de ined as he a io o modula ion ampli ude in ensi y o he pe cei ed signal by ha o he emi ed one: () pe emi m mF m = The m(F) alues a e ob ained o each modula ion equency F be ween 0.63 Hz and 12 Hz in 1/3 oc a e s eps (14 in o al). This ange co e s he speech modula ions. The ca ie noise signal is oc a e band il e ed be ween 125 Hz and 8 kHz. In his way he e a e 7×14=98 m(F) alues. These modula ion educ ion ac o s a e con e ed in o appa en signal/noise a ios and a e p ope ly weigh ed and a e aged o calcula e he STI index [3]. Fig.1 shows a possible scheme o calcula e MTF. This me hod is implemen ed, o example, in he ee PC-so wa e STI e .-2.0 om Lexing on School o he Dea [4], whe e he s imulus signal consis s o a band limi ed whi e ca ie noise modula ed by low equency sine wa es. Each measu e akes abou 2 minu es o apply all he modula ion equencies o he es sys em successi ely. Hougas and S eeneken [2] hemsel es showed ha o he majo i y o ac ual si ua ions in enclosu es, his se o 98 da a cons i u ed an unnecessa y de ailed g id o analysis, so hey de eloped a educed pa ame e : he RASTI index. The p ocess is e y simila bu he ca ie is il e ed o 500 and 2000 Hz only and he modula ion equencies a e also educed (4 o he 500 Hz oc a e band and 5 o he 2 kHz one). This measu ing me hod is implemen ed by B uel & Kjæ in he Speech T ansmission Me e ype 3361 [5] and by B üel Acous ics in he sys em RASTI 44BA wi h some new ea u es [6]. Sch oede [7] p oposed an al e na i e me hod o he scheme o Fig. 1 o ind he m(F) alues o linea , passi e and ime in a ian sys ems om he impulse esponse h( ). In his case, i ’s necessa y o il e h( ) o each oc a e band o he ca ie , . The m(F) alues a e calcula ed om he il e ed impulse esponse h ( ) acco ding o: 2 0 2 0 ()·exp(2) () () hjFd mF hd τπττ ττ ∞ ∞ − =∫ ∫ The impulse esponse can be ob ained using an impulsi e signal (sho ), bu in his case we canno ake in o accoun he backg ound noise because he spec um o he impulsi e signal is e y di e en om he no malized speech spec um de ined by Hou gas and S eneken [1]-[3], o e en hose p oposed by IEC [8]. In able 1 we show he no malized alues o he speech e e ence spec um a 250 Hz. This di icul y can be o e come using Maximum Leng h Sequences (MLS) as s imulus signal in o de o ob ain he impulse esponse, as Ri e poin ed ou [9]. In his case i ’s possible o inse a il e be ween he MLS gene a o and he inpu o he ampli ie o adap i s spec um. Table 1 also shows he il e ans e unc ion no malized o he alue a 250 Hz oc a e band and is compa ed wi h he e e ence speech spec um in Fig. 2. In o de o ob ain eliable MTF measu es, in addi ion o gene al equi emen s o measu es wi h MLS signals, we mus keep in mind some conside a ions, such as: • Measu ing an impulse esponse long enough (a leas 1 s) wi h a minimum bandwid h o 12 kHz (3 kHz o RASTI measu emen s). • No a e aging o ob ain he impulse esponse. This will a i icially educe he signal/noise a io ( he S/R inc ease +3 dB e e y ime he numbe o MLS pe iods a e doubled). E en hough i ’s possible o ake ad an age om his si ua ion o e alua e he incidence o backg ound noise agains o he modula ion deg ada ion ac o s. Table1.- IEC speech spec ums and il e ans e unc ion used o he MLS measu es. Oc a e Re . spec um Male oice Female oice Measu ed il e 125 -2,5 2,9 -200,0 -0,55 250 0,0 2,9 5,3 0,00 500 -1,0 -0,8 -1,9 -1,43 1000 -5,0 -6,8 -9,1 -4,60 2000 -10,0 -12,8 -15,8 -9,52 4000 -17,0 -18,8 -16,7 -16,80 8000 -23,0 -24,8 -18,0 -25,30 F equency (Hz) 125 250 500 1000 2000 4000 8000 A enua ion (dB) -30 -25 -20 -15 -10 -5 0 5 Re e ence speech Adap e il e Fig 2.- Re e ence speech spec um and il e ans e unc ion. Sys em unde es S imulus signal: Response signal x 2 x 2 1 2 (ca ie )(1sin(2)) m NoiseF π ×+ sys em noise & dis o ion]  +    emi mpe m Fig. 1.- Concep ual scheme o measu ing MTF using modula ed noise. • Rep oducing he MLS signal by an acous ic sou ce wi h he same di ec i i y as he human head has, i he e isn’ an elec oacous ic sys em. I i exis s, he signal will be applied o he audience by he loudspeake s o he sys em. • Reducing he bandwid h o a oid ime aliasing e o s when he e e be a ion ime is oo long (abou 7 s in ou sys em). Di iding by wo, he bandwid h allows us o double he e e be a ion ime o eliable measu es (al hough his implies ha he high equency bands a e no compu ed o calcula e he STI). EXPERIMENTAL RESULTS AND DISCUSSION We p esen he e he measu ed da a o h ee di e en ooms, which a e in he High School o A chi ec u e: he Physics Labo a o y, he Con e ence Hall and he Spo s Cen e. Thei app oxima e olumes a e 320, 1870 and mo e han 9300 m 3 espec i ely (see Fig. 3). Thei e e be a ion imes (T20 and EDT) o he oc a e bands be ween 125 and 8000 Hz we e measu ed using MLS echnique h ough he analyse based in PC MLSSA om DRA labo a o ies [10]. Fo each measu e we a e age eigh MLS pe iods in o de o imp o e he S/N a io. The signal was acqui ed by he omnidi ec ional mic ophone B&K-4269. These esul s appea in ig. 4. To e alua e he speech in elligibili y we ha e measu ed STI/RASTI indices using wo di e en sys ems: (i) he B üel & Kjæ Speech T ansmission Me e ype-3361 uses he modula ed noise echnique. The emission le el was adjus ed o Re +10 dB; his implies a SPL o 69 dB a he 500 Hz oc a e band and 60 dB a 2 kHz, measu ed a 1 m om he sou ce (≅67 dB(A)). The du a ion o each measu e is se up o 16 s. We ha e been using his echnique o se e al yea s wi h good esul s [11]. (ii) The MLSSA analyse uses he MLS echnique. The spec um en elope o his signal is desc ibed by sin2(x)/x2. The sampling equency is adjus ed by MLSSA o h ee imes he demanded bandwid h, so ha in ha condi ion he spec um is p ac ically la . Fo his eason he MLS signal is condi ioned by he il e CWF- 1 om One o One Technical P oduc s be o e en e ing he ampli ie inpu . The il e ans e unc ion was measu ed using MLSSA and is displayed in Fig. 2. The ampli ied signal is F equency (Hz) 125 250 500 1000 2000 4000 8000 Re e be a ion ime (s) 0 1 2 3 4 5 T20 EDT Con e ence Hall Spo s Cen e Labo a o y Fig. 4.- Measu ed e e be a ion imes. Labo a o y Con e ence Hall Spo s Cen e Fig. 3.- G ound plans o he measu ed ooms (no a he same scale). deli e ed o he audience by wo di e en sou ces, a dodecahed al sou ce B&K-4296 and a e e ence acous ic sou ce B&K-4205. The signal was adjus ed, ollowing manu ac u e ’s ecommenda ions, o p oduce a sound p essu e le el o 67 dB(A) a a sou ce-mic ophone dis ance o 1m. This le el was checked by he sound le el me e B&K-2231. In all cases he emi e was loca ed a he posi ion ma ked S in Fig. 3. The signal was cap u ed using wo di e en omnidi ec ional mic ophones: he B&K-4190, calib a ed be o e measu ing, and he AT4050/CM5 om Audio-Technica, which was unable o be calib a ed. Each one o hem wi h hei espec i e p eampli ie and pola iza ion sou ce (B&K- 2804 o he i s case and LAB1 om Ea hwo ks o he second one). In o de o e alua e he e ec o backg ound noise we made calib a ed measu es, wi hou a e aging, and uncalib a ed measu es, a e aging eigh MLS pe iods, o ob ain h( ). To calcula e STI alues i ’s possible o choose among h ee di e en spec ums: e e ence one, men spec um o women spec um om IEC-268-16 [8]. The alues we ha e used in his pape co espond o e e ence spec um. Fig. 5 shows RASTI alues measu ed in o de o e alua e he luc ua ions caused by small imp ecisions in he placemen o he mic ophone. Measu emen s we e pe o med bo h wi h he B&K (modula ed noise) and MLSSA analyse (MLS signal), o a ixed mic ophone posi ion (P-5 in labo a o y) and mo ing he mic ophone a ound his posi ion. The blue lines co espond o he 95% con idence in e al (±0.02) ha we can ake as andom e o s. As we can see, he e a e no meaning ul di e ences o ei he he wo measu emen echniques o o he mic ophone posi ions. The backg ound noise o he measu emen ime was s a iona y and lowe han 35 dB(A). Measu emen s o all posi ions in he labo a o y wi h he wo men ioned sou ces ha e been ca ied ou (Fig. 6). Some mino di e ences appea be ween he alues measu ed wi h di e en echniques (e en be ween RASTI and STI measu ed wi h MLSSA). The backg ound noise was simila o he p e ious si ua ion. Meaning ul dispe sions o he wo sou ces do no appea . These di e ences a e only no iceable when he mic ophone is loca ed la e ally om he sou ce due o he oice di ec i i y (see P-1, P-3 in Fig. 8). In o de o dispose o a g ea e in e al o STI/RASTI alues, we ha e epea ed hese measu emen s wi h inc easing wideband s a iona y backg ound noise le els p oduced by he e e ence acous ic sou ce B üel & Kjæ ype 4205 (Fig. 7). In his si ua ion he RASTI and STI alues measu ed wi h MLSSA appea di e en . We could expec his esul since RASTI only compu es he e ec s o 500 and 2000 Hz oc a e bands, while STI is a ec ed om wideband noise spec um. The di e ences be ween RASTI alues om B&K-3361 and hose calcula ed om impulse esponse measu ed wi h MLSSA a e s ongly dependen on he noise le el. The mayo di e ences appea (Fig. 7(b)) when he noise le el is simila o he signal le el, 64dB(A). Simila esul s a e ob ained in he Con e ence Hall (Fig. 8), measu ed wi h HVAC o (low and s a iona y noise le el: L eq=36 dB(A); L 10=37 dB(A) and L 90=35.5 dB(A)) and he HVAC on. In his case he backg ound noise e ec can be obse ed a posi ions a ound he g id (12-16) ( he measu ed noise le els a poin 16 a e Leq=49.8 dB(A); L10=50.5 dB(A) and L90=49.5 dB(A)). The Numbe o measu e 1 2 3 4 5 6 7 8 9 10 11 12 RASTI 0,0 0,2 0,4 0,6 0,8 1,0 B&K-3361 ixed B&K-3361 no ixed MLSSA (RASTI) ixed MLSSA (RASTI) no ixed 95% con idence Fig. 5.- Measu ed RASTI alues a P-5 (labo a o y) wi h he mic ophone a ixed posi ion (“ ixed”) and mo ing i a ound (“no ixed”). Measu emen posi ion 123456789 RASTI 0,0 0,2 0,4 0,6 0,8 1,0 Modula ed noise RASTI MLS signal (s. omni) RASTI MLS signal (s. B&K4205) STI MLS signal (s. omni) Fig. 6.- Measu ed RASTI alues a di e en posi ions in labo a o y wi h wo loudspeake s. di e ences a e mo e no iceable o he esul s om he Spo s Cen e (Fig. 9), whe e he a ic noise le el is highe and, abo e all, ime- a ian (Leq=47 dB(A); L10=50 dB(A) and L90=42 dB(A)). I backg ound noise is negligible and he e e be a ion e ec is he main eason o he loss o in elligibili y, hen STI alues de i ed om uncalib a ed MLS signals (a e aging se e al MLS cycles in o de o ob ain impulse esponse) a e simila o ha ob ained om calib a ed ones (wi hou a e aging) (Fig. 8, HVAC o ). We ha e also compa ed he esul s ob ained om wo e y di e en mic ophones wi h MLSSA. The esul s a e e y simila (Fig. 10). The eason o his compa ison is ha Audio-Technica is a mul ipa e n mic ophone which can be used o measu e o he acous ic pa ame e s (LF, LFC) simply selec ing be ween omnidi ec ional and igu e eigh pa e n. (a) Measu emen posi ion 123456789 STI / RASTI 0,0 0,2 0,4 0,6 0,8 1,0 Modula ed noise RASTI MLS signal STI MLS signal (b) Measu emen posi ion 123456789 STI / RASTI 0,0 0,2 0,4 0,6 0,8 1,0 (c) Measu emen posi ion 1 2 3 4 5 6 7 8 9 STI / RASTI 0,0 0,2 0,4 0,6 0,8 1,0 (d) Measu emen posi ion 1 2 3 4 5 6 7 8 9 STI / RASTI 0,0 0,2 0,4 0,6 0,8 1,0 Fig. 7.- Measu ed RASTI / STI alues wi h di e en wideband backg ound noise le els: (a) 54 dB(A), (b) 64 dB(A), (c) 74 dB(A) and (d) 80 dB(A); all o hem measu ed a 1m om noise sou ce. Measu emen posi ion 2 4 6 8 10 12 14 16 18 20 22 RASTI 0,0 0,2 0,4 0,6 0,8 1,0 Calib a ed MLS signal (HVAC o ) Calib a ed MLS signal (HVAC on) Uncalib a ed MLS signal (HVAC o ) Fig. 8.- RASTI alues o he Con e ence Hall measu ed a e aging 8 cycles MLS signal and wi hou a e aging (HVAC on and o ). Measu emen posi ion 2 4 6 8 10 12 14 16 18 20 22 STI / RASTI 0,0 0,2 0,4 0,6 0,8 1,0 Modula ed noise MLS signal (RASTI) MLS signal (STI) Fig. 9.- STI/RASTI measu ed a he Spo s Cen e. Finally, Fig. 9 ep esen s he co ela ion be ween all measu ed da a using he B&K and he MLSSA analyse wi h he 95% p edic ion in e al. In i , we ha e ma ked he quali ica ion in e als: g een o “bad” , pink o “poo ”, cyan o “ ai ”, blue o “good” and ed o “excellen ”. We can see how he quali ica ion de i ed om MLSSA alues can be di e en (and lowe ) om hose de i ed om B&K alues mainly in he “poo ” in e al. I would be necessa y o make mo e measu emen s o ob ain mo e alues in he ange good and excellen . CONCLUSIONS We ha e compa ed wo di e en echniques o measu ing speech in elligibili y. We ha e obse ed a mino in luence o he mic ophone and o he loudspeake in he MLS echnique. Howe e , he p esence o backg ound noise supplies alues ha di e om hose measu ed wi h noise modula ed echnique and, abo e all, i he backg ound noise is o he same o de o he es signal o ime- a ian . P obably in his si ua ion he condi ion o ime-in a ian i ’s no ul illed by he oom and consequen ly he in elligibili y quali ica ion can di e acco ding o he echnique used. We ha en’ ca ied ou he analysis o he measu ed MTF due o space easons. REFERENCES [1] S eeneken H.J.M. and Hougas T., The modula ion ans e unc ion in oom acous ics as a p edic o o speech in elligibili y, Acus ica 28, 66-73, 1973. [2] Hougas T. and S eeneken H.J.M., A mul i-language e alua ion o he RASTI-me hod o es ima ing speech in elligibili y in audi o ia, Acus ica 54, 185-199, 1984. [3] Hougas T. and S eeneken H. J. M., A e iew o he MTF concep in oom acous ics and i ’s use o es ima ing speech in elligibili y in audi o ia, J. Acous . Soc. Am. 67, 1060-1077, 1985. [4] h p://hea ing esea ch.o g/STI.h m [5] Speech ansmission me e ype 336. Re e ence manual. B üel&Kjæ , Næ um, Denma k. [6] h p://www.b uel-ac.com (E-mail: [email p o ec ed]) [7] Sch oede M. R., Modula ion ans e unc ion: de ini ion and measu emen , Acus ica 49, 179-182, 1981. [8] Red a o IEC 268-16, sound sys em equipmen , pa 16: The objec i e a ing o speech in elligibili y by he speech ansmission index. [9] Ri e D., Modula ion ans e measu emen wi h maximum leng h sequences, Jou nal o he Audio Enginee ing Socie y 40 (10), 779-790, 1992. [10] MLSSA, Maximum-leng h sequence sys em analyse , e e ence manual, .-10W. DRA labo a o ies. [11] Galindo M., Zama eño T., Gi ón S., Speech in elligibili y in Mudeja -Go hic chu ches. ACUSTICA Ac a Acus ica 86, 381-384, 2000. RASTI B&K 0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0 RASTI MLSSA 0,0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1,0 y = 1,061*x - 0,0986 2 = 0,883 Fig. 11.- Co ela ion be ween di e en echniques RASTI measu ed alues in he h ee ooms. Measu emen posi ion 2 4 6 8 10 12 14 16 18 20 22 STI 0,0 0,2 0,4 0,6 0,8 1,0 Mic ophone B üel & Kjae Mic ophone Audio-Technica Fi g. 10.- STI alues measu ed wi h wo di e en mic ophones a he Con e ence Hall.