comparison of long-term average speech

Transcription

1 An nternatonal spectra comparson of long-term average speech Dens Byrne, Harvey Dllon, and Khanh Tran Natonal Acoustc Laboratores, 126 Grevlle St., Chatswood, 2067 NSW, Australa Stg Arlnger Unversty Hosptal, Lnkopng, Sweden Keth Wlbraham Unversty of Manchester, Manchester, Unted Kngdom Robyn Cox Veterans Affars Medcal Center, Memphs, and Memphs State Unversty, Memphs, Tennessee Bjorn Hagerman Karolnska nsttute, Stockholm, Sweden Raymond Hetu GA UM, Montreal, Canada Joseph Ke and C. Lu Specal Educaton, Kowloon, Hong Kong Jurgen Kesslng Unversty Hosptal, Gessen, Germany M. Nasser Kotby, Nasser H. A. Nasser, and Wafaa A. H. El Kholy An Shams Unversty, Caro, Egypt Yasuko Nakansh a) Gakuge Unversty, Tokyo, Japan Herbert Oyer Oho State Unversty, Columbus, Oho Rchard Powell Taranak Hosptal, New Plymouth, New Zealand Dafydd Stephens, Rhys Meredth, and Tony Srmanna Welsh Hearng nsttute, Cardff, Wales George Tavartkladze and Gregory. Frolenkov Research Center for Audology, Moscow, Russa Soren Westerman and Carl Ludvgsen Wdex, Vaerloese, Denmark (Receved 4 November 1993; accepted for publcaton 1 July 1994) The long-term average speech spectrum (LTASS) and some dynamc characterstcs of speech were determned for 12 languages: Englsh (several dalects), Swedsh, Dansh, German, French (Canadan), Japanese, Cantonese, Mandarn, Russan, Welsh, Snghalese, and Vetnamese. The LTASS only was also measured for Arabc. Speech samples (18) were recorded, usng standardzed equpment and procedures, n 15 localtes for (usually) ten male and ten female talkers. All analyses were conducted at the Natonal Acoustc Laboratores, Sydney. The LTASS was smlar for all languages although there were many statstcally sgnfcant dfferences. Such dfferences were small and not always consstent for male and female samples of the same language. For one-thrd octave bands of speech, the maxmum short-term rms level was 10 db above the maxmum long-term rms level, consstent across languages and frequency. A "unversal" LTASS s suggested as beng applcable, across languages, for many purposes ncludng use n hearng ad prescrpton procedures and n the Artculaton ndex. PACS numbers: Gr, Ar, Ts a)currently at the Unversty of Tsukuba J. Acoust. Soc. Am. 96 (4), October /94/96(4)/2108/13/$ Redstrbuton subject to ASA lcense or copyrght; see Download to P: On: Thu, 29 Oct :43:15

2 NTRODUCTON Representatons of the long-term average spectrum of speech (LTASS) have varous acoustcal and audologcal applcatons. n rehabltatve audology the LTASS s wdely used n the prescrpton and evaluaton of hearng ad fttngs. t s used n many hearng ad prescrpton procedures ether n the dervaton of the prescrptve formula (e.g., Berger et al., 1977; Seewald et al., 1985; Byrne and Dllon, 1986) or n calculatng the prescrpton for the ndvdual clent (e.g., Cox, 1988; Sknner, 1988). The Artculaton ndex (ANS, 1969), whch has many current or potental uses ncludng the evaluaton of hearng ad fttngs (e.g., Pavlovc, 1989), depends on usng the LTASS. Although there are many publshed measurements of the LTASS, no partcular set of values s unversally accepted. ndeed, the dfferent hearng ad selecton procedures use varous sets of values (Sknner, 1988) and, although the Artculaton ndex Standard ncludes an dealzed LTASS, ths dealzed LTASS s not usually recommended n audologcal applcatons (Pavlovc, 1989). The majorty of publshed measurements of the LTASS are for the Englsh language, as spoken n the U.S.A. (Dunn and Whte, 1940; Stevens et al., 1947; Rudmose et al., 1958; Benson and Hrsh, 1953; Harrs and Wate, 1965; Nemoller et al., 1974; Pearsons et al., 1977; Cox and Moore, 1988; Cornelsse et al., 1991; Stelmachovcz et al., 1993), Australa (Byrne, 1977; Byrne and Dllon, 1986), England (Boothroyd, 1967), or other countres (Tarnoczy, 1971). There are also measurements for other languages whch nclude: German, Hungaran, talan, Russan (Tarnoczy, 1971), Swedsh (Anansson, 1974; Lejon, 1989), Dansh (Dalsgaard and Pedersen, 1966), Fnnsh (Kukaannem, 1980; Kukaannem et al., 1982), Mandarn (McCullough et al., 1993), French and Dutch (Harmegnes and Landercy, 1985), Polsh (Zalewsk and Majewsk, 1971), and Spansh (Banuls-Terol, 1971). All measurements of the LTASS are smlar but small dfferences occur between Englsh as spoken n dfferent countres and among dfferent languages. However, t s not clear whether these dfferences are real, because there are large dfferences among ndvduals (Dunn and Whte, 1940; Byrne, 1977; Kukaannem et al., 1982) and many of the studes have used only small subject groups. Furthermore, some studes (Tarnoczy and Fant, 1964; Dalsgaard and Pedersen, 1966; Tarnoczy, 1971; Nemoller et al., 1974; Zalewsk and Majewsk, 1971; Banuls-Terol, 1971) do not permt any estmate of ndvdual varablty because the analyses have been based on a "chorus" of all (or groups of) subjects combned. By comparng three studes that dd use substantal subject groups (at least 20), Cox and Moore (1988) concluded that there probably are small dfferences n the LTASS of Amercan and Australan speech. From examnng samples of Swedsh, Hungaran, and German, Tarnoczy and Fant(1964) concluded that there were sgnfcant dfferences among these languages n the mdfrequency regon ( Hz for males, Hz for females). They attrb- uted ths dfference to the relatve occurrence of vowels wth second formants n ths regon. Harmegnes and Landercy compared Dutch and French speech spectra of 20 male talkers who spoke both languages. They concluded that ndvdual talker dfferences accounted for most of the varablty n spectra but thathe 'e probably were small language df- ferences whch dd not exceed 5 db n any frequency regon. They suggested that the dfferences probably arose from dfferences n phoneme dstrbutons for the two languages. One partcular effect, whch occurred around 1000 Hz, was attrb- uted to the exstence of nasalzed vowels n French contrast- ng wth the absence of any such vowels n Dutch. On the other hand, from a comparson of the LTASS of French, Dutch, Englsh, talan, and Dansh (two male and two female talkers for each language), Pavlovc et al. (1991) concluded that there were no sgnfcant effects of ether sex or language. Overall, t appears that the LTASS may vary wth language but the ssue s stll not resolved. n addton to subject varaton, small dfferences could occur among dfferent sets of measurements because of varatons n measurement technques. For example, Dunn and Whte (1940) commented on some earler measurements that contaned an artfact, apparently due to close talkng condtons. n the varous studes there have been dfferences n the recordng or analyss condtons and these dfferences may be responsble for some small dfferences n results. One mportant procedural varable s the angle of ncdence of the recordng mcrophone to the talker's mouth. The majorty of studes have used 0 ø ncdence (.e., mcrophone drectly n front of mouth) recorded n anechoc or approxmately equvalent condtons. Some, however, have recorded a "chorus" of talkers wth the mcrophone placed n the dffuse far feld (Tarnoczy and Fant, 1964; Dalsgaard and Pedersen, 1966; Tarnoczy, 1971). Owng to the drectonalty of the human mouth/head and torso (Dunn and Farnsworth, 1939), the hgh frequences wll manly be radated n the frontal drecton whle the dffuse feld wll represent a spatal ntegraton of radaton n all drectons. Furthermore, most room surfaces have a greater absorptvty for hghfrequency sounds than for low-frequency sounds, so reverberant or dffuse sound felds wll tend to have relatvely weaker hgh-frequency components when compared to anechoc sound felds. Consequently, the relatve hghfrequency content wll be hgher n the frontal/anecho condtons. Nonetheless, the dfferences n LTASS between the chorus studes and most of the other studes are not large and do not affect wthn-study comparsons of languages. Notable dfferences may be obtaned when speech s recorded wth an angle of ncdence substantally dfferent from 0 ø (Studebaker, 1985). Cornelsset al. (1991) recorded speech at the ear of the talker and Stelmachovcz et al. (1993) recorded the speech of parents at the ear poston of a chldwth varous postural postons (e.g., sttng adjacent, hp, 2109 J. Acoust. Soc. Am., Vol. 96, No. 4, October 1994 Byrne et al: Long-term average speech spectra 2109 Redstrbuton subject to ASA lcense or copyrght; see Download to P: On: Thu, 29 Oct :43:15

3 cradle). These last two studes reflect (ntentonally) head dffracton and body baffle effects as well as varatons resultng from the drectonalty of speech. n the present study, the recordng mcrophone was placed at 45 ø ncdence to the talker's mouth axs and at a dstance of 20 cm. At ths close dstance, whch was selected to optmze the sgnal-to-nose rato and to mnmze the effect of any reverberant feld present, t was consdered undesrable to poston the mcrophone drectly n front of the mouth because of breath noses from plosve sounds. The use of 45 ø, compared wth 0 ø, has been shown to result n a relatve reducton of about 2 db at frequences from 1000 to 5000 Hz and slghtly more at 8000 Hz (Studebaker, 1985). Our choce of ncdence could be consdered to be a compromse between 0 ø, whch maxmzes the relatve hghfrequency content, and reverberant condtons or other angles, whch result n varyng amounts of reducton n hgh frequences. The LTASS wll undoubtably be nfluenced by the type of speech materal that s analyzed. For example, the accuracy wth whch the Artculaton ndex (A) predct speech recognton test scores wll be optmzed by usng the LTASS of the materal n that test (Studebaker and Sherbecoe, 1992). The LTASS of a lst of words or nonsense syllables may be very dfferent from the LTASS of runnng speech especally f the word or syllable lst contans frequent repetton of a few phonemes. (A strkng example s shown n Fg. 2 of Byrne, 1986 whch presents the LTASS of a nonsense syllable test.) Our nterest, however, n common wth that of the above cted authors, was not n the LTASS of any specfc materal but rather n dervng a LTASS that would be representatve of speech encountered over a range of everyday stuatons. t seems from the agreement of the above studes and from comparsons of dfferent materals (Benson and Hrsh, 1953) that the choce of materal s not crtcal provded that t s not grossly unrepresentatve phomencally, such as speech passages contanng repetton of a few phrases. Although the dfferences n measurements of the LTASS are small, t would be desrable to establsh a standard LTASS for everyday speech. t would be useful f such a LTASS could be taken as representatve of all, or a wde range of, languages for use n hearng ad prescrptve procedures. Some prescrpton procedures are used wth clents who speak languages (e.g., Cantonese) for whch LTASS measurements have not, to our knowledge, been publshed. There s, therefore, a practcal ssue n decdng whether such procedure should be used, wthout modfcaton, for clents who lsten to a language that dffers from the one for whch the procedure was developed. n a smlar ven, the A and other predctve procedures have been used wth varous languages although usually the users have made spectral measurements for the language concerned (Anansson, 1974; Lejon, 1989). The present study was desgned to examne the feasblty of developng a standard LTASS that would represent a TABLE. Speech sample nformaton. No. of Language Country talkers nvestgators Englsh England 32 Bamford, Wlbraham Australa 30 Byrne, Dllon, Tran New Zealand 21 Powell U. S. A., Memphs 22 Cox, Alexander U.S. A., Columbus 21 Oyer, Lambert Swedsh Sweden, Stockholm 22 Hagerman Sweden, Lnkopng 20 Arlnger Dansh Denmark 20 Westerman, Ludvgsen German Germany 27 Kesslng French Canada 20 Hetu Japanese Japan 27 Nakansh Cantonese Hong Kong 25 Ke Mandarn Hong Kong 21 Lu Russan Russa 21 Tavartkladze, Frolenkov Welsh Wales 23 Stephens, Meredth Snghalese Wales 21 Stephens, Srmanna Vetnamese Australa 19 Byrne, Dllon, Tran Arabc Egypt 20 Kotby, Nasser, E1 Kholy wde range of languages or, alternatvely, to dentfy any sgnfcant dfferences that may exst among languages. The research strategy was to record and analyze samples of a wde varety of languages and dalects, accordng to a standardzed protocol. Although t s establshed that the LTASS vares wth vocal effort (Tarnoczy, 1971; Pearsons et al., 1977), the present study was confned to "normal" vocal effort as t seems unlkely that varatons wth vocal effort would be language dependent. n addton to the LTASS, as defned by ts rms levels, t was of nterest to consder the dynamc range of speech across frequences. Ths was accomplshed by determnng, at three frequences, the levels exceeded varous percentages of tme. Ths type of nformaton, also presented by Dunn and Whte (1940), s relevanto predctng understandng of speech and to evaluatng the requrements of amplfcaton systems. Dynamc range metrcs, such as peak to rms dfferences, dffer accordng to speech materals (Studebaker and Sherbecoe, 1992) and other factors. t seems possble that dynamc range may also dffer across languages but we are not aware of any nformaton on ths pont. t s well establshed (Tarnoczy, 1971; Byrne, 1977; Pearsons et al., 1977; Cox and Moore, 1988) that the LTASS dffers for men and women. The man dfference s n the Hz frequency range and reflects the generally lower fundamental frequency range of male voces. Although ths dfference should apply across languages, and there s evdence that t does (Tarnoczy, 1971), there may be other more subtle dfferences that could be language or dalect dependent. The present study, therefore, ncludes separate analyses of data for male and female voces.. METHODS A. Overvew Eghteen speech samples were recorded n 13 countres (15 locatons) usng standard sets of equpment and a stan J. Acoust. Soc. Am., Vol. 96, No. 4, October 1994 Byrne et al: Long-term average speech spectra 2110 Redstrbuton subject to ASA lcense or copyrght; see Download to P: On: Thu, 29 Oct :43:15

4 dard protocol. The recordngs were then analyzed at the Natonal Acoustc Laboratores (Australa). The recordngs represented 13 languages ncludng Englsh as spoken n four countres. Table lsts the languages and countres represented, the numbers of talkers n each sample, and the nvestgators w ho made each recordng. Most languages were recorded n the countres of ther orgn, the exceptons beng Snghalese and Vetnamese whch were recorded n Wales and Australa, respectvely. B. Talkers Wth one excepton, each recordng contaned at least ten male and ten female talkers aged between 15 and 60 years. (There were only nne female Vetnamese talkers.) All talkers spoke the language concerned as ther frst language and none had any obvous speech defects. No other selecton crtera were employed. C. Recordng equpment Eght sets of recordng equpment were assembled. These conssted of a hgh qualty (but nexpensve) cassette tape recorder deck (Techncs RS-B105) and a custom-made mcrophone unt based on a Knowles EA 1934 mnature mcrophone. A set of equpment and a tape was sent to each nvestgator and was returned wth the recordng. Durng analyss, each recordng was replayed on the equpment on whch t had been recorded. The frequency response of each recordng and playback system was relatvely flat and correctons were made for the mnor dscrepances that exsted (see Sec. F). D. Speech materal A passage was selected from a story book on the bass that t was relatvely easy to read and dd not nvolve excessve repetton. (Most materal would meet ths latter crteron except some nursery rhymes.) Ths "standard" passage was used for all recordngs of Englsh but other materal, meetng the same crtera, were used for other languages. The materal took about 90 s for most talkers to read and provded more than the requred 64 s of speech for all talk- ers. E. Recordng procedure Whenever possble, recordngs were made n an anechoc room at least 3 m wde by 3 m long by 2 m hgh. However, ths crteron was relaxed n several nstances because such a faclty was unavalable. The talker read the materal whch was enlarged and placed on a chart at least 1 m n front of hm or her. Thus the talker was able to look straght ahead throughouthe recordng and the procedure avoded any possblty of reflectons from materal held n the talker's hands. The recordng mcrophone was on a stand or suspended (.e., not on a table or other reflectng surface) n front of the talker, 20 cm from and n the same horzontal plane as the mouth and at an azmuth of 45 ø ncdence, rela- tve to the axs of the mouth. The talker was nstructed to read aloud at a normal speed and level. Before recordng, each talker read the passage slently and then read t aloud at least once to ensure reasonable fluency. The talker was nstructed that t dd not matter f there were some mstakes and that he or she should keep on readng rather than stoppng to correct any errors. The practce readng was also used to set the tape recorder to an approprate recordng level (adequate but wth mnmal overloadng) and ths settng was noted so that absolute levels could be calculated later. (Although not calbrated, the recorder volume control was large and well marked and thereby permtted settngs to be reproduced accurately.) F. Analyss procedures Each recordng of speech was analyzed, usng a Bruel & Kjaer 2131 analyzer coupled to a Tektroncs computer, to derve overall and thrd-octave band rms levels, averaged over 64 s of sgnal. Absolute levels were derved by comparng the output ntensty to that of a standard pure-tone cal- brator whch was used to record an 84 db SPL tone on the tape and was replayed. The frequency response of each mcrophone and tape recorder was analyzed at the center of each 1/3 octave frequency from 100 to Hz. As the varaton between the correcton fgures for each set of equpment relatve to the average correcton fgure was less than 1.2 db, and for most frequences was less than 0.5 db, a common set of correcton fgures was used for all recordngs. Values for 63, 80, , and Hz were obtaned by extrapolaton, and the results at these frequences should thus be treated wth some cauton. (n fact, the LTASS values for 63 Hz are questonable because they may also be nfluenced by nose.) The largest correcton fgure used was +7.5 db, for Hz. Wthn the range Hz, no correcton fgure exceeded 2.0 db. The dynamc range of the speech was assessed by the followng procedure. Ten talkers (fve female and fve male) from each sample were randomly chosen for analyss. The tape recorder output was nput to four Bruel & Kjaer 2231 sound level meters (SLMs), whch had nstalled the DZ7101 statstcs module. One of the sound level meters measured the broadband sgnal and the other three were connected to Bruel & Kjaer /3 octave flter sets. These were set to 400, 1000, and 4000 Hz. The SLMs were set to detect the rms envelope of the nput sgnal usng a "Fast" tme constant (conformng to EC 651). Ths tme constant s nomnally 125 ms. The SLMs were programmed to measure for 64 s and the followng parameters were recorded from each of the four sound level meters at the end of the analyss perod. "Peak" refers to the hghest nstantaneousgnal level measured durng each passage. "Max" refers to the hghest level of the rms envelope measured durng each passage. "L eq refers to the long-term rms value durng the entre analyss tme. "L" refers to the envelope value exceeded 1% of the tme. L10, L50, L90, and L99, smlarly refer to the envelope values exceeded 10%, 50%, 90%, and 99% of the tme, respectvely. The SLM sampled the envelope at ntervals of ms J. Acoust. Soc. Am., Vol. 96, No. 4, October 1994 Byrne et al.' Long-term average speech spectra 2111 Redstrbuton subject to ASA lcense or copyrght; see Download to P: On: Thu, 29 Oct :43:15

5 G. Statstcal treatment The long-term equvalent 1/3 octave and overall levels obtaned from the spectrum analyss were processed n two ways. The frst concentrated on examnng dfferences n spectral shape, whle the second concentrated on examnng dfferences n overall level. One-thrd octave levels for each talker were corrected by the measured frequency response of the tape recorder and mcrophone. These values were normalzed so that the longterm overall (lnear frequency weghtng) level was 70 db. Two-way analyss of varance (ANOVA) was performed wth "Sample" as a between-groups varable and "Frequency" as a wthn-groups varable. ("Sample" s used n preference to "country,""language," or "locaton" because, n some nstances, two languages were recorded n the same locaton and, n other nstances, the same language was recorded n two or more countres or locatons.) Separate analyses were performed for the male and female talkers. To enable a ready comparson among samples, the grand average spectrum was calculated by averagng, separately for males and females, the mean spectrum for all samples excludng Arabc (see later). The mean for each sample, males and females separately, was then plotted aganst the grand average spectrum. The sgnfcance of devatons from the grand average spectrum was assessed as follows. The grand average spectral levels were subtracted from the normalzed 1/3 octave levels at each frequency for each talker. At each frequency, a t test was then used to compare the mean dfference for each sample (Arabc excluded) wth the null hypothess of zero dfference. Because there were 850 such tests (25 frequences by 17 samples by two sexes), we would expect 42 to be sgnfcant by chance alone f the usual sgnfcance level of 0.05 was adopted. For ths reason, Fgs. 1-4 show only whch devatons from the grand average are "sgnfcant" wth p<0.01. The overall levels n db SPL wth a flat frequency weghtng were analyzed n a two-way ANOVA wth sex and sample as between-group varables. H. Analyss of Arabc The Arabc speech, whch became avalable later than the other samples, receved a more lmted analyss. The LTASS was determned by the procedures descrbed above but dynamc range measurements were not undertaken. Arabc was not ncluded n the statstcal treatment (ANOVAs) or n determnng the "unversal" LTASS to be presented later. The sgnfcance of dfferences between Arabc and the unversal LTASS was assessed by t tests.. RESULTS A. Shape of the speech spectrum For both the male and female talkers, the major effects of sample and frequency, and the nteracton between, them were all sgnfcant wth p < Fgures 1-4 show the male and the female LTASS for each sample, except Arabc. Also shown s the average LTASS for all samples combned. Separate male and female values are shown for frequences below 200 Hz but, for the rest of the frequency range, the average s for males and females combned. Symbols shown n bold are those devatons from the average that are sgnfcant at the 0.01 level and flled symbols show devatons that are sgnfcant at the level. Consderng Fgs. 1-4, t s clear that the number of "sgnfcant" devatons (204) s far more than would be expected by chance alone (eght). n vew of the numerous comparsons, we suggesthat devatons should only be consdered to be truly sgnfcant when they are at the level, or when devatons at the 0.01 level occur n two or more adjacent frequency bands. The values for the grand average spectra (ncluded n Fgs. 1-4) for males, and females, are shown n Table. The thrd column shows a combned long-term average spectrum. For frequences up to and ncludng 160 Hz, t s equal to the male spectrum. For hgher frequences, t s equal to the average of the values for males and females. As wll be dscussed later, ths s recommended as an approprate unversal spectrum for many, but not all, purposes. B. Overall levels The ANOVA (wth non-normalzed data) showed that sample was hghly sgnfcant (p < ) but that sex was not sgnfcant at the 0.05 level. The average value for males was 71.8 db SPL, whle that for females was 71.5 db SPL. The average values for each sample (both sexes combned) ranged from 67.8 for Vetnamese to 75.2 db for Man- darn. The dstrbuton of overall levels for all talkers s shown n Fg. 5. C. Dynamc range All values for each talker were normalzed by subtractng tha talker's Leq value for the respectve band. The resultng relatve levels are shown n Fg. 6. The only major varatons between the samples occurred for the L90 and L99 percentle levels. These may not be ndcatve of genune dfferences between the languages and dalects because these levels are presumably affected by the amount of background nose present durng pauses n the contnuous dscourse. A four-factor ANOVA was performed on the data wth the L 90 and L99 data excluded. Between-groups factor were sample and sex, and repeated measures factors were band and percentle. The man effects are of no nterest, and of the 11 nteracton effects the followng were sgnfcant: sex x band, sample x percentle, sex X percentle, band X percentle, sample X sex X percentle, sample X band x percentle, sex x band x percentle, and sample x sex x band x percentle. Wth the excepton of sample x sex x percentle (p=0.008), all of these were sgnfcant wth p< Despte the extremely hgh level of sgnfcance, the effects were generally only a few decbels n magntude, and the statstcal sgnfcance arose from the large number of talkers and observatons. Fgure 7 shows all the percentle levels as a functon of band and sex. D. ndvdual talker dfferences The spectra for ndvdual talkers showed substantal varatons. Ths was true for all samples although there was 2112 J. Acoust. Soc. Am., Vol. 96, No. 4, October 1994 Byrne et al: Long-term average speech spectra 2112 Redstrbuton subject to ASA lcense or copyrght; see Download to P: On: Thu, 29 Oct :43:15

6 65 : : AUSTRALA ' : ' NEW ZEALAND ' rlr,, 60 w 55 m 50 > 45 o ao., 35 c o 0, Q A T '''... r $5 6o, A USA MEMPHS ' ' ' r... ' ' ' ' ss so s o T ' ' j J a * UNTED, (/) z, ' KNGDOM ' lk 2k 4k 8k 16k Frequency (Hz) 'fl "' s, Z, õ o LEGEND O MALE FEMALE AVERAGE 0 HALE p < O. Ol FEMALE p < 0.0 MALE p < o. OOl o 30,, lk 2k 4k 8k Sk Fr'equency (Hz) FEMALE p < FG. 1. Male and female long-term average speech spectrum (LTASS) values for fve samples of Englsh. Sold lne shows LTASS average across 17 speech samples (all samples except Arabc), males and femaleseparately for frequences below 160 Hz, combned for hgher frequences. no statstcal examnaton of whether varablty nteracted wth sample. For all samples combned, Fg. 8 shows the standard devaton of ndvdual varatons from the mean value at each frequency for males and for females. The analyss s based on data whch had been normalzed to a 70 db overall level for each talker. The devaton values are smlar for both sexes and at all frequences from 630 to 4000 Hz. Varablty shows a small but consstent ncrease, 2113 J. Acoust. Soc. Am., Vol. 96, No. 4, October 1994 Byrne et al.' Long-term average speech spectra 2113 Redstrbuton subject to ASA lcense or copyrght; see Download to P: On: Thu, 29 Oct :43:15

7 ,, 41C SUEDS H,: n--.l ' ' LZNKOPZNG '',-, :o'...:---',---:- ': ': : $$ '--...,... m SO,-- --,...,---r---,- 45 ' øu 40 ' '... ß... 0,,,, 65 so v : : o : : ozsh : 55 "' ,--,...,---r---,- :-.,... ( 30 ' ' ',-, $5 o.. RUSSRN ,-,... r e... d..., 25; 250 5;00 Ftequenc l LEGEND m '... '... L... 0 M LE m m 50 A FEMRLE > 45 o 40.e &...e RVERRGE 0 M::LE p < O. Ol FEMRLE p < O. Ol 35 c o 30 e... d J.. J e......,...,...,.. f Frequenc l (Hz) HqLE p < o. OOl FEMRLE p < O. 00l FG. 2. Male and female LTASS values for Swedsh (two samples), Dansh, German, and Russan. Sold lne shows LTASS average across 17 speech samples. for both sexes, for bands above 4000 Hz. The only substantal ncreases n varablty are n the bands 80 and 100 Hz for males and 125 and 160 Hz for females. E. LTASS for Arabc The LTASS values (db) for Arabc, normalzed to 70 db overall level (lnear), are shown n Fg. 9.. DSCUSSON The overall fndng of ths study s that the LTASS s very smlar over the wde range of languages that were analyzed. ndeed, there s no sngle language or group of languages whch could be regarded as beng markedly dfferent from the others. Therefore, t s feasble to propose a unver J. Acoust. Soc. Am., Vol. 96, No. 4, October 1994 Byrnet al.' Long-term average speech spectra 2114 Redstrbuton subject to ASA lcense or copyrght; see Download to P: On: Thu, 29 Oct :43:15

8 ,-, $5... O ' ' CANTONESE Q) SS ' - ' ' T E 0 o 0 ' ' -... T... r o $0,., 65._1 t3_ (/) & , ss so > o S00 lk 2k 4k 8k lsk, $5 Frequenc j (Hz) so LEGEND ' 55 S0 O MALE / FEMALE -, AVERAGE O MALE p < 0. 0], FEMALE p < O. O l 40 O MALE p < m m m m m ' m A FEMALE p < O. OOl 030 m,, m m m ß m,, m, ]k 2k 4k 8k ]6k Frequenc l (Hz) FG. 3. Male and female LTASS values for Cantonese, Mandarn, Vetnamese, Japanese, and French. Sold lne shows LTASS average across 17 speech samples. sal LTASS that would be applcable to most (possbly all) languages and would be suffcently precse for many purposes. Nonetheless, there are small but (statstcally) sgnfcant dfferences among languages and more substantal dfferences, at the low frequences, between male and female talkers. A. Male/female dfferences Consderng frst the comparson between males and females, the most notable feature s that ther spectra are vrtually dentcal over the frequency range from 250 to 5000 Hz. Wthn ths range, the normalzed male and female lev J. Acoust. Soc. Am., Vol. 96, No. 4, October 1994 Byrne et al.' Long-term average speech spectra 2115 Redstrbuton subject to ASA lcense or copyrght; see Download to P: On: Thu, 29 Oct :43:15

9 ,,,.,.....,.,,., 65 SNGHALESE 60 ' ' - - ' ss m so e, e, e 5 o o 0.c 35.,...,...'...,...'...,...,. o S00 lk 2k 4k 8k $k Frequency (Hz) LEGEND S0 500 lk.2k 4k 8k lsk Frequenc l (Hz) MLLE /, FEMRLE MRLE p < 0.01, FEMRLE p < 0.01 MRLE p < , FEMRLE p < RVERRGE FG. 4. Male and female LTASS values for Snghalese and Welsh. Sold lne shows LTASS average across 17 speech samples. TABLE. Male, Female, and combned speech spectra, normalzed for 70 db SPL overallevel and averaged acrossamples (excludng Arabc). Combned s equal to male for frequences up to 160 Hz; t s average of male and females for other frequences. The spectrum (combned male and female) recommended by Cox and Moore (1988) s also shown. Frequency Cox and Moore (Hz) Male Female Combned (1988) els, averaged over all languages, agree wthn 2 db at all thrd-octave frequences except 800 Hz, where the dfference s 2.3 db. (The non-normalzed values agree almost as closely as there was lttle dfference between the average overall male and female speech levels.) For frequences of 160 Hz and below, male levels greatly exceeded female levels undoubtably because of the dfference n the fundamental frequency ranges. These fndngs are consstent across languages and consstent wth prevous research (Benson and Hrsh, 1953; Tarnoczy and Fant, 1964; Tarnoczy, 1971; Nemoller et al., 1974; Byrne, 1977; Pearsons et al., 1977; Cox and Moore, 1988) m Long term rms level (db SPL) FG. 5. Dstrbuton of overall rms levels (measured at 20 cm from mouth). Curve shows normal dstrbuton ftted to data J. Acoust. Soc. Am., Vol. 96, No. 4, October 1994 Byrne et al.' Long-term average speech spectra 2116 Redstrbuton subject to ASA lcense or copyrght; see Download to P: On: Thu, 29 Oct :43:15

10 lo. Max... ß,)...,, ,,"-,a'- '.." ß ', ;.:.' ;'';.:e'" ' " ''', -''"''-.:,'' '...,':... a,, ß o--, o-,,-o....o-,,-o --o ß -1:)-.)-- o -,-o. ß 'o o. N -r ;''-' '6.= ' ';'. ' :' ' ' 7;,'"' %'; "- ' 7: ' ' :' ' ' "-: ':'. ' :':' ': '. '"'- ' '1. ' 0 O -10 > o o a.:.,ao,.O....,a'--,D,a--a..,o..a., ',a""- 190 ß " ß...:'"o'-'.-o'"'...'%0"...,o,,,.....' , - L, 9 ' (c) > > -- "'T' Otm.-- Z 0. V N -- 'r (D "' Z -r' ß -J (D (,0 H Z LU <1 <1 Z D Z Z 0. (D r)' <1 (_1Z -J D _.J 'r' <1 H (. -) H < "' D LJ n O H a -a-.a -'>' -a.-a- -a- a- -.o- a. --a- o..-a- a -a.,. *a '"a L10 o.. -o,-'.o---o.-o ß ß a -- a. ",a.. a,' ' o a..-o- ß -a...o--.a-,-a.,t- 150 F OOF Z N œ Z H Z* Z D ZZ ½ OZ DJ ½HU hh œ D,O OH Stud Stud FG. 6. (a) Peak, maxmum, and percentle levels of the broadband sgnal (relatve to Leq ) for each sample. Peak, maxmum, and percentle levels of the (b) 400, (c) 1000, and (d) 4000 Hz 4001/3 octave bands (relatve to the Leq for that band) for each sample. For 6300 Hz and hgher frequences, female levels exceeded male levels. Ths was a consstent fndng; when averaged across the frequences from 6300 to Hz nclusve, the average level for females exceeded that for males for every country although only margnally so n some nstances. Ths s shown n Fgs The average dfference between males and females, n ths range, was 2.6 db, wth d, fferences rangng from 0.3 to 5.9 db. The same trend occurred n the Arabc data (Fg. 9) n that the female values exceeded the male values by an average of 2.9 db. Some prevous studes have found a hgher overall level, typcally 2-3 db, for male than for female voces (Benson and Hrsh, 1953; Byrne, 1977; Pearsons et al., 1977). Our data do not show a sgnfcant dfference although the small ß C}' "- -0 o. ß 43 a---a O'. --- ß '0 a---a... a O... a-e a 0... o... 0.'.'.' a a L n Sex: Band: F M F M F M F M Broad 400 Hz khz 4 khz Frequency (Hz) FG. 7. Percentle levels (relatve to band L eq ) versus talker sex and mea- surement band. FG. 8. ndvdual varablty n speech levels averaged across 17 samples (all samples except Arabc) normalzed for L eq' 2117 J. Acoust. Soc. Am., Vol. 96, No. 4, October 1994 Byrne et al.' Long-term average speech spectra 2117 Redstrbuton subject to ASA lcense or copyrght; see Download to P: On: Thu, 29 Oct :43:15

11 F'.. ' ' ^ ' RVERRGE -' " ' g... :: :: :: c 0 S 0... t J- - -'-... t...,, t, G lk 2k 4k 8k lgk Frequenc (Hz) FG. 9. Male and female long-term average speech spectrum for Arabc. Sold lne shows LTASS averaged across the other 17 samples (.e., excludng Arabc). mean dfference that exsts (0.3 db) s n the predcte drecton. Pavlovc et al. (1991) have also reported no sgnfcant dfference for sex. The dscrepancy n the fndngs of dfferent studes may be nfluenced by dfferences n vocal effort as Pearsons et al. found that the male/female dfference n overall level ncreased wth ncreasng vocal effort. Also, any ntrnsc male/female dfferences could possbly be reduced by nstructons f they dscouraged the use of partcularly soft or loud voces. There appears to be some male/female dfference wth respect to the degree of ndvdual talker varablty n the lower frequency bands. Ths affects manly the frequency bands somewhat below the typcal male and female fundamental frequences and s probably related to dfferences n the fundamental frequency range of ndvdual talkers. B. Language/dalect dfferences ndvdual languages showed many statstcally sgnfcant varatons from the average values. However, most of the varatons are less than 3 db n magntude, or occur outsde the range of 200 to 6300 Hz, or occur only n one solated 1/3 octave band. f we examne only those statstcally sgnfcant varatons whch are larger than 3 db, and whch occupy at least two adjacent bands wthn the range 200 to 6300 Hz nclusve, we fnd only the followng varatons from a "unversal speech spectrum:" (1) New Zealand male and female speech s hgh by 3-6 db from 5000 Hz and above; (2) Vetnamese male and female speech s low by 5-6 db from 5000 Hz and above; (3) Arabc male and female speech s low by 5 db at frequences around 5000 Hz; (4) Japanese male speech s low by 3 db around 2500 Hz; (5) Cantonese male speech s hgh by 3-4 db from 630 to 1250 Hz; (6) Australan female speech s hgh by 3-4 db from 5000 Hz and above; (7) Memphs female speech s hgh by 3-4 db from 1250 to 2000 Hz; (8) Russan female speech s hgh by 3-4 db at 250 and 315 Hz; (9) Mandarn female speech s hgh by 3-4 db from 2500 to 4000 Hz; and (10) Snghalese female speech s low by 3-5 db from 4000 Hz and above. Explanatons for the above varatons are not obvous. t may be that some or all of them occur because dfferent languages or dalects use somewhat dfferent vowels or use the same vowels or other sounds, such as "s," wth dfferent frequences of occurrence. Prevous authors have sometmes suggested explanatons of ths type for the dfferences they found among the LTASS of dfferent languages (Tarnoczy and Fant, 1964; Harmegnes and Landercy, 1985) or dalects (Cox and Moore, 1988). However, our data show that when languages devated from the average, they often dd so for only one sex. Therefore, any factors whch explan the devatons must nteract wth sex dfferences. The dffculty of fndng credble explanatons for dfferences n the LTASS may be llustrated by comparng the fve samples of Englsh (Fg. 1) wth respecto sex dfferences. n the frequency range above 5000 Hz, male/female level dfferences occur for three samples but not for the other two samples. t s dffcult to understand how ths sex dfference could occur n Australan speech but not n New Zealand speech, or how t could occur n one Amercan sample but not n the other. t s clear that t would be a major undertakng to attempt to relate dfferences n languages or dalects to dfferences n the LTASS. Such an nvestgaton was beyond the scope of the present study. An obvous queston s whether some or all of the small dfferences found could be explaned by dfferences n the recordng technques used at dfferent locatons. Aganst ths explanaton s the pont just mentoned, namely that dfferences from the LTASS often occurred for only one of the sexes. Furthermore, varatons tend to occur over several adjacent thrd-octave bands. t therefore seems unlkely that the varatons are caused by room resonances, whch tend to be more localzed n frequency. We note also that n the nstances where two languages were recorded n the same localty, they do not show smlar devatons. Thus t seems hghly unlkely that dfferences n recordng technque could explan the dfferences noted above. (Possble exceptons could be the New Zealand and Arabc samples as both sexes show the same trend and there are no other recordngs from the same localty for comparson.) From consderng the above noted varatons and others occurrng at frequences below 200 or above 6300 Hz, there appears to be no systematc separaton between the Englsh versus the non-englsh languages, or between the nontonal versus the tonal languages (Cantonese, Mandarn, Vetnamese). Ths last fndng agrees wth that of McCullough et al. (1993) who found no dfference between the LTASS of Englsh and Mandarn d. Acoust. Soc. Am., Vol. 96, No. 4, October 1994 Byrne et al: Long-term average speech spectra 2118 Redstrbuton subject to ASA lcense or copyrght; see Download to P: On: Thu, 29 Oct :43:15

12 C. Effects of normalzaton/analyss methods Most of the above devatons occur n the hghfrequency or mdfrequency regon. However, the nature of the devatons observed wll depend partly on the choce of analyss method. The normalzaton process ensured that the overall level of all talkers was set to 70 db SPL. Because the overall level s domnated by the more ntense low-frequency bands, there s less chance of observng dfferences among talkers n ths regon. The alternatve of normalzng at one specfc frequency would cause the same problem to a greater degree. The use of non-normalzed levels would make the detecton of dfferences n shape dependent on the range of overall levels allowed by each expermenter because varaton of overall level would add to the varance among talkers at each frequency. The sgnfcance of the major effect of "sample" wthn the analyss of varance of spectral shape s at frst surprsng because the data had been normalzed so that all talkers had an overall level of 70 db SPL. The effect s readly explaned by varatons n the hgh-frequency regon. Because the speech spectrum s weghted towards the low frequences, the overall level, based on a power addton of the ndvdual bands, s lttle affected by the level of the lower ntensty, hgh-frequency components. The ANOVA statstc, however, sums the decbel value of all bands lnearly, so varatons n the hgh-frequency level affect the major effect of sample just as much as do varatons n the low-frequency levels. D. Dynamc range The dynamc range measurements, lke the LTASS measurements, show that all languages are smlar despte there beng a number of statstcally sgnfcant, but small, dfferences. Essentally, the L 50, L 10, L 1, Max, and Peak values are equvalent across all samples (Fg. 6) and for males and females (Fg. 7). There s one dfference between our data and those of other studes n that our data show a 10-dB dfference between the L 1 and L values n contrast to the eq wdely accepted value of 12 db (ANS, 1969), although only slghtly dfferent from the value of 11 db shown by the data of Cox et al. (1988). Another mnor dscrepancy that our data do not show the L 1/L eq dfference to ncrease wth fre- quency, as s shown by some other data (e.g., Fletcher and Galt, 1950). E. Overall level The measured overall speech levels of our subjects averaged about 72 db SPL for both male and female talkers. As the recordng mcrophone was only 20 cm from the mouth, ths level would correspond to 58 db SPL at a dstance of 1 m, assumng no reverberaton. Ths level s about 5 db less than s usually reported for conversatonal speech (Pearsons et al., 1977). We suggesthat, despte nstructons to speak "normally," many talkers tend to speak at a low level when a mcrophone s placed close to them. Ths may be analogous to the natural tendency to adjust voce levels accordng to the dstance from lsteners. n smlar ven, t has been shown repeatedly (see Byrne, 1983 for revew) that the typcal speech nput receved by moderately mpared hearng ad wearers s about 70 db SPL whch requres greater than normal vocal effort. We beleve that no sgnfcance should be attached to the overall levels found n a study such as ours except to note that dfferences n vocal effort affect the LTASS. Greater vocal effort wll result n an ncrease n the relatve md- to hgh-frequency content of speech, although the dfferences between soft and average speech are small (Pearsons et al., 1977; Kukaannem et al., 1982). F. Unversal LTASS The smlarty of the LTASS across samples demonstrates that t s reasonable to propose a unversal LTASS whch should be satsfactory for many purposes and applcable to most, f not all, languages. Such a LTASS should be sutable for hearng ad prescrpton procedures and for the Artculaton ndex, regardless of language. The comparablty of speech dynamcs across languages also supports the dea that the applcaton of such procedures to dfferent languages s not complcated by acoustcal dfferences n conversatonal speech. Of course, there may well be other complcatons, such as possble dfferences n frequency mportance functons across languages, whch would need to be nvestgated before applyng procedures lke the Artculaton ndex unversally. Our recommended unversal LTASS s very smlar to the recommendaton of Cox and Moore (1988)(see Table ) but t has the advantages of havng been derved from a larger total data set and of representng a range of languages. t may also be more realstc because t s based on a 45 ø rather than 0 ø angle of ncdence. As mentoned earler, substantally dfferent speech spectra have been obtaned for measurements made at dfferent angles of ncdence wth respect to the talker and when nfluenced by varatons n head/body dffracton effects (Cornelsse et al., 1991; Stelmachovcz et al., 1993). Those measurements were prompted by an nterest n prescrbng amplfcaton for chldren and the recognton that speech s often presented from drectons other than drectly n front of the chld. n a revsed verson of a hearng ad prescrpton procedure, Seewald (1992) opted to use a compromse between the usual (0 ø ncdence) LTASS and one recorded at the ear of the talker. The possble merts of such a choce wll not be consdered here but t s mentoned to show that there may be an argument for usng dfferent LTASS values for partcular purposes. Nonetheless, our unversal LTASS could serve as a bass from whch any requred varatons could be made. There could be crcumstances where separate male and female spectra would be desrable, namely applcatons where the very low frequences are sgnfcant. However, a sngle LTASS s suffcent and, therefore, preferable for the applcatons consdered here, that s, n relaton to hearng ad prescrpton and the A or smlar procedures for predctng speech ntellgblty. Our unversal LTASS should also be suffcently precse for a range of more general applcatons concernng the desgn or use of speech transmsson systems. Fnally, t s clear that the LTASS and the dynamc characterstcs of conversatonal speech are very much domnated by the characterstcs of the vocal mechansm. Although dfferent languages use dfferent vowels (formant structures) 2119 J. Acoust. Soc. Am., Vol. 96, No. 4, October 1994 Byrne et al.: Long-term average speech spectra 2119 Redstrbuton subject to ASA lcense or copyrght; see Download to P: On: Thu, 29 Oct :43:15

13 and the frequency of occurrence of varous phonemes dffers, these factors appear to have had only mnor effects on the LTASS and dynamc measures of conversatonal speech. Furthermore, these dfferences are not even consstent across male and female talkers of the same language. ACKNOWLEDGMENTS n addton to the authors, the followng partcpated n the study: John Bamford (Unversty of Manchester), Geneveve Alexander (VA Medcal Center, Memphs), and C. Lambert (Oho State Unversty). We also thank Dr. Margaret W. Sknner for helpful comments on a draft of ths manuscrpt. Anansson, G. (1974)."Speech dscrmnaton predcted from tone audometry and artculaton ndex," Acta Oto-Laryngol. Suppl. 320, ANS (1969). ANS S , "Amercan Natonal Standard Methods for Calculaton of the Artculaton ndex" (Amercan Natonal Standards nsttute, New York). Banuls-Terol, V. (1971)."Weghted average spectrum of human speech: An approach," Proceedngs of the 7th CA, Budapest, pp Benson, R. W., and Hrsh,. J. (1953)."Some varables n audo spectrometry," J. Acoust. Soc Am. 25, Berger, K., Hagberg, N. S., and Rane, R. L. (1977). Prescrpton of Hearng Ads (Herald, Oho). Boothroyd, A. (1967)."The dscrmnaton by partally hearng chldren of frequency dstorted speech,"nt. Audol. 6, Byrne, D. (1977)."The speech spectrum--some aspects of ts sgnfcance for hearng ad selecton and evaluaton," Br. J. Audol. 11, Byrne, D. (1983)."Theoretcal prescrptve approaches to selectng the gan and frequency response of a hearng ad," Monogr. Contemporary Audol. 4, (1), Byrnes, D. (1986)."Effects of frequency response characterstcs on speech dscrmnaton and perceved ntellgblty and pleasantness of speech for hearng-mpared lsteners," J. Acoust. Soc. Am. 80, Byrne, D., and Dllon, H. (1986). "The Natonal Acoustc Laboratores' (NAL) new procedure for selectng the gan and frequency response of a hearng ad," Ear Hear. 7, Dalsgaard, S.C., and Pedersen, S. B. (1966)."The power densty spectrum of Dansh speech," Transactons of Dansh Engneerng, No , 1-7. Cornelsse, L. E., Gagn, J.P., and Seewald, R. C. (1991)."Ear level recordngs of the long-term average spectrum of speech," Ear Hear. 12, Cox, R. (1988)."The MSU hearng ad prescrpton procedure," Hear. nstrum. 39, Cox, R. M., Matesch, J. S., and Moore, J. N. (1988). "Dstrbuton of short-term rms levels n conversatonal speech," J. Acoust. Soc. Am. 84, Cox, R. M., and Moore, J. N. (1988)."Composte speech spectrum for hearng ad gan prescrptons," J. Speech Hear. Res. 31, Dunn, H. K., and Farnsworth, D. W. (1939)."Exploraton of pressure feld around the human head durng speech," J. Acoust. Soc. Am. 10, Dunn, H. K., and Whte, S. D. (1940). "Statstcal measurements on conversatonal speech," J. Acoust. Soc. Am. 11, Fletcher, H., and Galt, R. H. (1950)."The percepton of speech and ts relaton to telephony,"j. Acoust. Soc. Am. 22, Harmegnes, B., and Landercy, A. (1985)."Language features n the longterm average spectrum," Rev. Phon6t. Appl , Harrs, C. M., and Wate, W. M. (1965)."Measurements of speech spectra recorded wth a close-talkng mcrophone," J. Acoust. Soc. Am. 37, Kukaannem, H., Soponen, P., and Mattla, P. (1982). "ndvdual dfferences n the long-term speech spectrum," Fola Phonatr. 34, Kukaannem, H. (1980)."Speech ntellgblty hearng losses lnearly slopng to hgh frequences," Acta Unv. Oul. D 57. Opthalmol. Oto-rhnolaryngol Lejon, A. (1989). Optmzaton of hearng-ad gan and frequency response for cochlear hearng losses, Techncal Rep. No. 189, Chalmers Unversty of Technology, Gothenburg, Sweden. McCullough, J. A., Tu, C., and Lew, H. L. (1993)."Speech-spectrum analyss of Mandarn: mplcatons for hearng-ad fttngs n a mult-ethnc socety," J. Am. Acad. Audol. 4, Nemoller, A. E, McCormck,., and Mller, J. D. (1974)."On the spectrum of spoken Englsh," J. Acoust. Soc. Am. 55, 461. Pavlovc, C. V. (1989)."Speech spectrum consderatons and speech ntellgblty predctons n hearng ad evaluatons," j. Speech Hear. Dsord. 54, 3-8. Pavlovc, C. V., Ross, M., and Espesser, R. (1991)."Perceved spectral energy dstrbutons for EUROM.O speech and for some synthetc speech," Proceedngs of the X nternatonal Congress on Phonetc Scences, 5/5, Pearsons, K. S., Bennett, R. L., and Fdell, S. (1977). Speech levels n varous nose envronment, EPA Rep. No. 600/ Envronmental Protecton Agency, Washngton DC. Rudmose, H. W., Clark, K. C., Carlson, F. D., Esensten, J. C., and Walker, R. A. (1958). "Voce measurements wth an audo-spectrometer," J. Acoust. Soc. Am. 20, Seewald, R. C. (1992)."The desred sensaton level method for fttng chldren: Verson 3.0," Hear. J. 45, Seewald, R. C., Ross, M., and Spro, M. K. (1985)."Selectng amplfcaton characterstcs for young hearng-mpared chldren," Ear Hear. 6, Sknner, M. W. (1988). Hearng Ad Evaluaton (Prentce-Hall, Englewood Clffs, NJ). Stelmachowcz, P. G., Mace, A. L., Kopun, J. G., and Carney, E. (1993). "Long-term and short-term characterstcs of speech: mplcatons for hearng ad selecton for young chldren," J. Speech Hear. Res. 36, Stevens, S.S., Egan, J.P., and Mller, G. A. (1947)."Methods of measurng speech spectra," J. Acoust. Soc. Am. 19, Studebaker, G. A. (1985). "Drectvty of the human vocal source n the horzontal plane," Ear Hear. 6, Studebaker, G. A., and Sherbecoe, R. L. (1992)."A model for the predcton of average speech recognton performance of normal-hearng and hearng-mpared persons," Lab. Rep. No , Hearng Scences Lab, Memphs State Unversty. Tarnoczy, von T. (1971). "Das durchschnttlche energe-spektrum der sprache (fur sechsprachen)," Acustca 24, Tarnoczy, T., and Fant, G. (1964)."Some remarks on the average speech spectrum," Q. P.S. R. Rep. No. 4, pp , Speech Transmsson Laboratory, Stockholm. Zalewsk, J., and Majewsk, W. (1971)."Polsh speech spectrum obtaned from supermposed samples and ts comparson wth spectra of other languages," Proceedngs of the 7th CA, Budapest, pp J. Acoust. Soc. Am., Vol. 96, No. 4, October 1994 Byrne et al.: Long-term average speech spectra 2120 Redstrbuton subject to ASA lcense or copyrght; see Download to P: On: Thu, 29 Oct :43:15