{ "paper_id": "J77-1003", "header": { "generated_with": "S2ORC 1.0.0", "date_generated": "2023-01-19T02:49:53.516043Z" }, "title": "The Hearsa -I Speech Understanding System: An Example ot the J Recognition rocess", "authors": [ { "first": "I", "middle": [ ") Raj" ], "last": "Reddy", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Richard", "middle": [ "D" ], "last": "Fcaeetl", "suffix": "", "affiliation": {}, "email": "" }, { "first": "D", "middle": [ "C" ], "last": "Washi~ngfon", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Ricltard", "middle": [ "B" ], "last": "Ncelcy Xerox I'alo'", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Alio", "middle": [], "last": "Research Cerrter", "suffix": "", "affiliation": {}, "email": "" }, { "first": "J", "middle": [ "12 K" ], "last": "Ilni~l", "suffix": "", "affiliation": {}, "email": "" }, { "first": "P", "middle": [ "S" ], "last": "Llakcr", "suffix": "", "affiliation": {}, "email": "" }, { "first": "N", "middle": [ "R" ], "last": "Cobcii", "suffix": "", "affiliation": {}, "email": "" }, { "first": "!", "middle": [ "V" ], "last": "Dison", "suffix": "", "affiliation": {}, "email": "" }, { "first": "I", "middle": [ "?" ], "last": "Donald", "suffix": "", "affiliation": {}, "email": "" }, { "first": "", "middle": [], "last": "Walkcr", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Joseph", "middle": [], "last": "Fvcizenbaunr", "suffix": "", "affiliation": {}, "email": "" }, { "first": "W", "middle": [ "H" ], "last": "Freeman", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Conlpdny", "middle": [], "last": "Son", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Fra", "middle": [ "/" ], "last": "Lcisco", "suffix": "", "affiliation": {}, "email": "" }, { "first": "", "middle": [], "last": "Cullfornia", "suffix": "", "affiliation": {}, "email": "" }, { "first": "", "middle": [], "last": "1976", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Nu", "middle": [ "F" ], "last": "Iorial", "suffix": "", "affiliation": {}, "email": "" }, { "first": "S", "middle": [ "A Barchcnko" ], "last": "Ussr", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Alvin", "middle": [ "M" ], "last": "Libcrnlan", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Raimo", "middle": [], "last": "Rakis", "suffix": "", "affiliation": {}, "email": "" }, { "first": "F", "middle": [], "last": "Silvcrnia~l", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Sergei", "middle": [], "last": "Ivanov", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Y", "middle": [], "last": "Matsushita", "suffix": "", "affiliation": {}, "email": "" }, { "first": "H", "middle": [], "last": "Yamazaki", "suffix": "", "affiliation": {}, "email": "" }, { "first": "F", "middle": [], "last": "Sa'to", "suffix": "", "affiliation": {}, "email": "" }, { "first": "U", "middle": [], "last": "Warotan~akikkhadit", "suffix": "", "affiliation": {}, "email": "" }, { "first": "N", "middle": [], "last": "Kanchanawan", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Georgc", "middle": [ "W" ], "last": "Reinrherr", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Aaministrative", "middle": [], "last": "Correspondence", "suffix": "", "affiliation": {}, "email": "" }, { "first": "David", "middle": [ "E" ], "last": "Johnson", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Eva-Maria", "middle": [ "Mi" ], "last": "Wotscbke", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Detlef", "middle": [], "last": "Wotscbke", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Peter", "middle": [ "J" ], "last": "Downey", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Paltonio", "middle": [], "last": "Naun Fraya", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Aravind", "middle": [ "S" ], "last": "Joshi", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Ralph", "middle": [ "M" ], "last": "Weischedel", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Azriel", "middle": [], "last": "Rosen", "suffix": "", "affiliation": {}, "email": "" }, { "first": "Carl", "middle": [], "last": "Weiman", "suffix": "", "affiliation": {}, "email": "" } ], "year": "", "venue": null, "identifiers": {}, "abstract": "In l4enrsay-.I ciivcrsc sources of kt~owledpc can be represented as cooj~ernting independcat par:~llel processes which help i n the decodil~g of the utterssces using lllc hypotl~esize-r~ld-test pur;ldigol. , 'the system is discussed by collsiderillg a specific esn~iiple of its ol~eratiotl in the realm of voice chess. ~Gpics: feature extraction and sqnle~lt:ition, the recognition process, speaker-and environment-dcpaldent knowledge. sy~itnctic nnd semantic knowl6dge.", "pdf_parse": { "paper_id": "J77-1003", "_pdf_hash": "", "abstract": [ { "text": "In l4enrsay-.I ciivcrsc sources of kt~owledpc can be represented as cooj~ernting independcat par:~llel processes which help i n the decodil~g of the utterssces using lllc hypotl~esize-r~ld-test pur;ldigol. , 'the system is discussed by collsiderillg a specific esn~iiple of its ol~eratiotl in the realm of voice chess. ~Gpics: feature extraction and sqnle~lt:ition, the recognition process, speaker-and environment-dcpaldent knowledge. sy~itnctic nnd semantic knowl6dge.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Abstract", "sec_num": null } ], "body_text": [ { "text": "Ileccnt progress i n speech research is rev~ewed w~t h an cmphnsls on tlie efforts to relate ltrigilrsttc u111 ts to 5pecch cvcrl ts Topics disc i r f~i : Scgnlc~ltnl rlspects, As~in~rl;~tton, Syllables as cuiicateriatlve segment? (111stcad 01' photlenies). Supr:~srgctncntaI a3pects. In the d~sci~sston of sitpraseg~iis~~tals expcrlrnerlts are rel)ortt.d i n ~~h l~h E t i~l~\\ t i words were first recordcd in a camel scr~tencc and then const~tuent syllables were clipped out and used to forrii new sentences. Varrous approximatlo~ic; to thc target seiiterlcc weie tested. sinlplc concatenation, a ver5lon In which durations of syllables had been adju\\ltcl to m:itch nittilr,il cxpanslon or c.otnpres~ion of spllahlcs. anothcr versloll in whtc.11 the nalirral prk t~ contorti (taken ft-on^ an utterance of the tnlget bentencc) IS strctclicd to match the duratioi~s ' 1s they were for the source syllables, arid another version wlth 11ati1raI ~i t c l l colltour and dirrat~ori-r~djusted syllables.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "PSYCHOLOGY -9 2", "sec_num": null }, { "text": "A digital formulation of the plinse vocoder, an atialysis-syntlies~s systcm providing a parametric representation of a speech wavefarm by its short-time Fourier transform, is of interest both far data-rate reduction and for manipulatlng speech pnrnnieters. Thc system is des~gried to be an iduntity systeln 111 the absence of any parameter modifications. Computational eff~cret~cy is achreved by en~ploylng the fast Folirter tr;~nsforrn nlgorithm to perform the bulk of the coniptrtcltion in both the ari:~lysis and sy~~thcsis procedures, thereby milking the f o r~n i i l a t~o i~ attract~ve t'or irilplenientnt~on on a mirlico~i~puter.", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Departrnetlt of Electrical Enginreritlg and Cor?rplrf er Science, Research Laboratory of Electronics, M.I.T., Ctrmbridge, M A 02139 IEEE Transcictions orr Acoustics, S p e~c h , artd Sigrral I'rocessirrg ad: 243-248 Jltrte I976", "sec_num": null }, { "text": "Ry deco~nposing English syllables iri to phonetical'l y and phonotncticall y wcll-motivated units it should be possible to create a complete invuntury for st.gmeiitiil concatenation wliich will cont;lin about 1,000 entries, and still rcproducc 1iatitra1' a l l o p h o~i i~ yari3tio1is. A syllabic is shorn to consist of a syllablc core and syll;~ble affix(es), the former being decomposed into the' irli t i tal and final dcmisyllablc.~. Consonantal features t'or each demi-syllabe inclrrdcat most. one specification of place of articulation ( i n terms of i 1 few distinctive features), spi ran tiqtibn (for /sp/./st/,/sk/, as opposed to /p/. /t/, / k/). t e r i s~e s s . rlas;tl i ty etc.. ' 'The phonetic realization ir; governed hy thc vowel affinity principle as described clscwhere iii ternis of the te~iiporal scquencc of pcrtiner~t physical events. Tl~crt. is ;I pliorlctic corlstrsint that a deniisyll:tble is rca1i~t.d wltt! ~i o t rilore than ~wo',phoilctic colisona~ital s c~n~e n t s .", "cite_spans": [], "ref_spans": [], "eq_spans": [], "section": "Fujinlura Bell Laborntories, rllrrrrcay H111, Ncbv Jersey 07974 Pupcr Pres~nted nr ttle vrsr nleerrng of' the Acorrstical Society of Anlericu, April 1976", "sec_num": "0." }, { "text": "The final ~onsol~antal cle~nerlts (such ;is /s/ in /tacks/) tli;~t follow :I place-specified coiisotiiltit ( k i n tliis C;~SC) 3re tre;ltc'il ' ;IS syl1at)lt. nff iscs. Tl~ese affises ;Ire ;ill r~pic~nl, arid they observe voicirlg a s s i~i~i l a t i o~~ wi tli rcspcct to thc true consorl:rrlt i l l the core. 194-196 April 1976 I'he schenie uses ~n f o r m a t~o n about interphontrnie contextual effects coiimioed it1 fornlarit lransbt~ons and c~liploys internr~l trial sytithesis and feedbacl ~;\\nlpi\\risoiI (7s a 111ea1'ls for rccog~i~tion. The all11 is to achieve mintrnal \\c'i~slt~vlty tr) al>prc~i:~ol$ vurii~l>ility which occurs i n the speech signal, even for utterances of' a sit~gle speaker. Whrle the approach is quite gelrt.r;rl, it lias inrtially been tr~ed out on vowcl-stop-vohfel uftera~ices. Vowels are always ident~f~ed correctly wflrle rt'cognltlon scores range from 66% to 78% for the consotiarils, exccpt for / n / at 47.9%.", "cite_spans": [ { "start": 312, "end": 330, "text": "194-196 April 1976", "ref_id": null } ], "ref_spans": [], "eq_spans": [], "section": "Fujinlura Bell Laborntories, rllrrrrcay H111, Ncbv Jersey 07974 Pupcr Pres~nted nr ttle vrsr nleerrng of' the Acorrstical Society of Anlericu, April 1976", "sec_num": "0." }, { "text": "GLODlS represents a flexible, operattng-system approach to the gencratiorr and implcmentat~on of conlplex rules for deciston making i n p.attcrti recogrlition, such ;is processing of EEG. EI: k* Road, Hamburg, New York 14075-3c *", "num": null, "html": null, "text": "are reducing their budget of periodical subscriptions Many journals are no longer available to us * Members who subscribe to any journals other than the * * standard few can help the Association by providing -EC * is needed The member who wishes to help should send * a list of available journals to AJCL, 5048 Lake Shore * * Authors of technical reports can help by sending their articles, or summaries of them, to AJCL for publication.Many summaries ard authors' abstracts, gometimes edited for clarity, brevity, or completeness The bibliography is produced with programs designed by Martin Kay, using the equipment of the Xerox Palo Alto Research CenterAJCL gratefully acknowledges their contributionSee frame 58 for a list of subject headings with frame numbers" }, "TABREF1": { "type_str": "table", "content": "
Transition networks slid Chomsky's hierrcrchy: modified Eorley's algorithm for transition network grammars. Stochastic and error correcting versions of transition nerworks are proposed to solve the probleln of noise and distortion in syntactic palter11 recognition. This approach is ill ustrilted by discussion of an expcri men t on sorce-chess language. Inference in tfiinsition networks; ~nference on the probability assignment over the hrcs of s2ochastic tratisitioti networks; examples of inference.
PHONETICS-PHONOLOGY: RECOGNITION
", "num": null, "html": null, "text": "" }, "TABREF2": { "type_str": "table", "content": "", "num": null, "html": null, "text": "" }, "TABREF3": { "type_str": "table", "content": "
Report PX-/ 1534, 52p, 14 Novelrlber I975 NTIS: AD-A0 19 047/OGA
PC $4.5O/hlF $2.25
Computer prograins for detecting syntactic boundaries (ROUND3) and locating stressed
percs~vtxi bo~~nriaries wtsre evidenced by long tlilie ~I I~L ' I V ; \\~S ~) C~W C C I I syllables. The interstrcss i~ltcrval also providcd il good nle;\\suru of rate of spccch, that coirelatcd w~t h error rates i l l
i\\utom3tic plio~~etic classif t c a t i o~~ scllemes.
PHONETICS-PHONOLOGY: RECOGNITION: SPEAKER IDENTIFICATION
", "num": null, "html": null, "text": "hnve bee11 supplicd to ARl'A contractors arid incorporilted into speech recogrlition fnuilitrcs. Exper~ments w r e conductetl on vi~rious t~n l i r~g cues that correlatr! wrth photlol~gici~l and syntactic pllrase bouliclaries, sllowit~g that 91% of the phonological phrase boutldnr~es that were perceived by llsteilers who tlcard spectrally iiiverted speech could be detected frorn lengthcr~ed vowels and sonorall ts in p h~ ase-f ilia1 p o s~ t~ons. Also, 95% of these on an acoustic analysis of a fiucd, sentence-long utterance resliltlng it1 a firfiction of titne or coiltour for each fcaturc analyzed (such as pitch, ititensity). In a test of the systern 104 111rllc arid fe~r~ale spenhcrs, caIIed 111 nt11i1111311y once each worklr~g day, froni their owti phones, over a perlon of ftvc months. In the ~nitial call each 'customer' was asked to prov~dc 5 rccord~~igs of the test ~~tlcrr~nce (\"We were away a year ago\"). P r o v~s~o n ic made for updnt~ng the reference f~l e . At tile early stagcs the ryect-customer rat? is about 10%, while it i~~~p r o i l~h~~ 4% on adapted cttbtomers. I he nccept-~a~postcr 'rate shows a similar h~story. 7 hc greatest weakness of thc systenl 11t.s I n the estnbl~sh~ner~t of adccluate initial referehc~ ~les.PHONETICS-PHONOLOGY: RECOGNITION:Recent experiments in speech synthesis have shown that, by an appropriate eigenvector analysis, r set of ortliogonrl pnranlrters can be obtained that is essentially inde ~endeiit of all" }, "TABREF4": { "type_str": "table", "content": "", "num": null, "html": null, "text": "" }, "TABREF5": { "type_str": "table", "content": "
", "num": null, "html": null, "text": "hslowtl strings TI . . . 7%. \\tot. wrsk to cl~ousl: tliu ),lio\\vn string Ttrr" }, "TABREF6": { "type_str": "table", "content": "
d i~~u s s e d : 1) maximunl direction cosine tnetl~od, 2PHONETICS-PHONOLOGY: RECOGNITION CLASSIFICATION 1 1 1 1 I I I -I I I t I IIIICIC'F I i t I th;lt ' ttie nieasu~ cd paramctcrs are ~I~L~I I~I I ( c J ; I C C O I~I I I~ ( O ttlc 11111Ilid11iie11~101ial G;LIISSI~II prolx~bil~ ty dcnslty ft'~nct~o~?. 7 hc 111i*311\\ :tlid cotilr1;IIiCcs fur flit' C~;~~I S S I ; I I~ d~s t r~b~~f~o~i are (Ic~L'I I I I I I I C '~ l'ro111 ~l~i \\ l~l~; \\ l f y cli\\sslf'~ed 4pct'~ Residual
", "num": null, "html": null, "text": ") mtninrum dlstnnce with 110 mean correctrotl. 3) tninimum distance with linear Incan correction. 4) m~nimum distance with full correction. These were tested with 80, 40, 20. and 10-poltit spectral represeatation. Mcasures of accuracy and stahllity were derived through the use of an nilt~inatic performance evaluntlon system. Ovel 3000 hand-labeled spectra were uscd. Of those evaluated, a l i r~early mean-corrected minim utn distance measure, on a 40 pqlnt spectral representat1011 with a square (or cube) hot111 was consistently superior to the other! rnelliods. -unvoiccd decislor~ to pitch anr\\tysis r'\\ot only rcsults in unnecessary ~~~t i p l c \\~t y . blrt 111~tl.c~ i t difiicillt to classify short speech seglnctits which are less t11a11 n few pitch per~ods 111 tlumtioll. By using rneist~rcnlents of tllr xcro-crossi~~g.. ralc, the speech energy. corrcl;it~oll I~ctwocn aclj,lccnt speech snml~lcs. the first prcd~ctor cocff lclent fro111 a 12-pol2 111ic;ir I I~C~I C~I V L ' C ' U J I I I~ (LIT) ;\\~i:iIysis. ~11ii1 the energy in thc prcdlct~on error it is po\\\\~I)lc Lo usc pi\\tIcr~l recc)piiit~ori approach to dccidlng whctlier a seg1ne111 IS voiced speech ui~vo~ccci spccch. OI silcl~ce. The spcech sognlcnt is assrgrictl tc) n pitrticirlar clilss bltscd oil a t~ diltrl IIICILICIIIIS ; I lri111111ig be[. A s111lp1e rlolll irle,lr sn1oot11111~ algo~ 11hr)l is dcscr~bccf tu pro1 lctc : r srnnoi1i 3-levcl con totl~ of an uttornr~ce f o~ nsc i n \\1)cecb\\l rccogtli t l o l l :lppl I C~~~O I I S . Acoustics, Speech, and Signal Processing 24:" }, "TABREF7": { "type_str": "table", "content": "", "num": null, "html": null, "text": "" }, "TABREF8": { "type_str": "table", "content": "
Physiological measurements are mnde directly 011 human talkers to determine several dynamic laryngeal functions. The functions are c o~~t r o l
", "num": null, "html": null, "text": "" }, "TABREF9": { "type_str": "table", "content": "", "num": null, "html": null, "text": "" }, "TABREF10": { "type_str": "table", "content": "
", "num": null, "html": null, "text": "" }, "TABREF11": { "type_str": "table", "content": "
", "num": null, "html": null, "text": "" } } } }