Kassler 1972

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

Review: Optical Character-Recognition of Printed Music: A Review of Two Dissertations

Author(s): Michael Kassler


Review by: Michael Kassler
Source: Perspectives of New Music, Vol. 11, No. 1, Tenth Anniversary Issue (Autumn - Winter,
1972), pp. 250-254
Published by: Perspectives of New Music
Stable URL: http://www.jstor.org/stable/832471
Accessed: 31-05-2015 16:48 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact support@jstor.org.

Perspectives of New Music is collaborating with JSTOR to digitize, preserve and extend access to Perspectives of New Music.

http://www.jstor.org

This content downloaded from 128.192.114.19 on Sun, 31 May 2015 16:48:36 UTC
All use subject to JSTOR Terms and Conditions
OPTICAL CHARACTER-RECOGNITION OF PRINTED MUSIC:
A REVIEW OF TWO DISSERTATIONS

AUTOMATIC RECOGNITION OF SHEET MUSIC. By Dennis Howard


Pruslin.Sc. D. Dissertation,MassachusettsInstituteof Technology, 1966.

COMPUTER PATTERN RECOGNITION OF STANDARD ENGRA VED


MUSIC NOTA TION. By David StewartPrerau.Ph. D. Dissertation,Massachu-
settsInstituteof Technology,1970.

Readers of Perspectivesscarcelyneed be remindedof the pre-eminenceof


the written-musical domain (i. e., that domain of musical experiencein which
music is presentedvisuallyin one or another systemof musical notation) in
musicology:beforeEdison composerscould notproduce recordsof theirwork
in the sounded-musical domain, and other domains of musical experience
such as the tactile domain utilized in the Braille systemhave been employed
comparativelyinfrequently;and even afterEdison variousextra-musicalcon-
siderations (such as copyrightlaw and the relativelyhigh cost of sound-
processingmachinery)have joined with traditionto keep the written-musical
domain a principal mode of non-transientmusical communication.Within
this domain varioussystemsof musical notationhave achievedvariousdegrees
of currencyat various places and times, but of all these systemsone-the
currentcommon musical notation ('CCMN' for short)-has dominated: vir-
tually all music printedhas been printedin one or another'dialect' of CCMN:
even music originallynoted in another systemgenerallyhas been transcribed
into CCMN beforeprinting.
In recent years digital computers have become more efficientand more
prevalent,so that today, at least in computationallywell-developedparts of
the world, it no longer is unreasonable to delegate, or to plan to delegate,
musical processes to electroniccomputingmachinery.Of course,many musi-
cal processes do not involvepreviouslyrecorded musical compositions: per-
haps it is to the comparativelyearlysuccess of a few such computer-mediated
processesthatan unfortunatesynecdochicmisidentification of 'computermu-
sic' with 'synthesizingsound throughthe use of a digitalcomputer' has aris-
en.1 But (and of thistoo readerswill be well informed)centralto musicology
are processes that do involve prior musical compositions,and for the full
delegation of these processes to computingmachinerythe relevantcomposi-
tions must be put into computer-acceptableform.Human key-puncherscan
transcribefrom CCMN onto (say) punch cards (at PrincetonUniversitythe
Masses of Josquin were so transcribed,at a rate of approximately20 minutes
per printedCCMN page), but as this task clearlyrequiresno intelligencebe-
yond that with which machines can be endued it is only natural to consider

1See,forinstance,
p. 477 ofthe1971-72calendaroftheUniversity
ofCalgary.

S250

This content downloaded from 128.192.114.19 on Sun, 31 May 2015 16:48:36 UTC
All use subject to JSTOR Terms and Conditions
COLLOQUY AND REVIEW

constructingan artifactthat can scan printed CCMN pages, recognize and


discriminateamongst the 'primitivesymbols' of CCMN there arrayed,and
finallynotifythe resultof thisrecognitionin computer-acceptableform.2Al-
though optical character-recognition (OCR) machinesthat 'read' printedtext
(in Roman and Cyrillicalphabets) have been operationalforseveralyears,the
dissertationsunder review report-so far as I am aware-the firstand only
substantial endeavours to apply OCR technologyto the 'reading' of printed
CCMN. Both authors have essayed to solve less than the entireproblem, so
the work of each should be judged by its extensibilityto, ratherthan by its
non-realizationof, an actual workingmachine.
Dr. Pruslinchose to limit attention to a subsystemof CCMN in which
music is noted in just one measureon only two parallelstaffs.Chords sharing
a common stem are allowed on each staffto consist of up to four notes, of
which no two adjacent notes are more than two staff-positions apart; but 'the
sometimes occurringcase in which two parallel melody lines exist for the
righthand, left hand, or both, is not handled' (p. 18). The subsystemlacks
rests, accidentals, half-notes,whole-notes,and notes attached to more than
one stem; eighth-notesand notes of lesser note-durationare admittedonly if
beamed to an adjacent note. Clefs, time-signatures, grace-notes,dynamic
marks,phrase marks,and other 'special signs' of CCMN are disallowed.
Pruslin'srecognitionscheme begins with a scanningdevice that convertsa
measure of music (noted in this subsystem) to a facsimilearray of about
65,000 black-or-white pointsthatare storedstraightforwardly in a digitalcom-
puter. A computerprogramprocesses this arrayto sharpenthe focus both of
horizontal lines (which may come to be recognized as staff-linesor beams)
and of verticallines (which may come to be recognizedas stemsor bar-lines);
when in proper focus the staffsare located and the widths of the stafflines
are computed.
Pruslinthen isolates 'contour traces' by subtractingelectronicallyboth the
staff-linesand the verticallines and segmentingthe residue so that-roughly
speaking-withincertainsize limitationseach contourtrace encloses a minimal
area that contains adjacent black points but is bordered entirelyby white
points. Each contour trace is potentiallyeithera 'note cluster',i. e., a note or
a dialectallypermittedchord of two to fournotes, or one or more beams used
to connect eighth-notesor notes of lesser note-duration.A large portion of
the dissertationis devoted to detailed presentationof the particularalgorithms
used to classifycontour traces.
Afterrecognitionof the note clustersand the beams, a comparativelysim-
ple procedure assignsnote-durationalvalues (e. g., 1/4) to the note clusters.
Pruslindoes not go on to compute the pitchesthat the notationindicates,but

2Somerequirements thatsucha devicewillhaveto fulfill


arestatedin my'Anessay
towardspecification of a music-reading
machine'(Princeton, 1963,mimeographed;re-
printed1971 by the AmericanMusicological Society(GreaterNewYork Chapter)on
pp. 151ff.of theirill-edited
volumeMusicology and theComputer of
(CityUniversity
NewYorkPress)).

S251

This content downloaded from 128.192.114.19 on Sun, 31 May 2015 16:48:36 UTC
All use subject to JSTOR Terms and Conditions
PERSPECTIVESOF NEWMUSIC

fromhis results(obtained usingthe TX-0 and IBM 7094 computers)the staff-


positions of notes are readily determinable,and fromthis stage other algo-
rithmscan take over.3
Pruslin'sreportsof his experimentsare encouraging.'The note clusterrec-
ognition procedures were tested using a total of 153 note clusters.Withthe
criticalvalue of template match set at 70%, only one error,the rejectionof a
note cluster occurred' (pp. 88-89). Similar success is reportedfor other rec-
ognitionproceduresconstitutingPruslin'sscheme.However,one mustbalance
this encouragementagainst the small size of the experimentalsample (which
appears to have been selected fromjust one musicalpublication) and the ab-
sence of evidence that Pruslin'srecognitionscheme providesa sensible basis
forthe ultimatelywanted machinethat recognizesall of CCMN.
Indeed, Dr. Prerauat firstplanned merelyto extend Pruslin'smethodsto
cover the parts of CCMN that Pruslinhad excluded.4 But upon finding'that
Pruslin's preprocessingerased or distortedmost music symbols other than
quarter-notesand beamed notegroups,the objects that his programwas de-
signedto recognize' (p. 42), Prerauchose a differentapproach. Afterscanning
and digitizingtwo to threemeasuresof a musical score, his computerprogram
locates 'fragments'-contiguousblack regionsthat are presenteitherbetween
a pair of adjacent staff-linesor just above or just below a staff.Next these
'fragments'are 'assembled', according to rules that connect certainadjacent
fragments, into 'components' eitherto isolatedmu-
sic symbols(e. g., an accidental that--roughly--correspond
or a note not beamed to any adjacent note) or
to a group of beamed notes. The recognitionof components as one or more
particularmusic symbols is a multi-stageprocess, involvingsize ('a naturalis
a little tallerthan a flat' (p. 120)) as well as syntax('a sharpcan appear .., .as
an accidental, to the left of a note . . [or] in a key signature,to the rightof
a clef or a bar-line'(p. 144)).
Preraulimitsattentionto a subsystemof CCMN in which each of only two
parallel staffsbears monolynear5music composed of notes, rests,treble and
bass clefs, certain time-signatures, 'sofons'6, and dots of prolongation-but
not tempo indications, dynamic or phrase marks, or certain other 'special
signs'. This CCMN dialect is instancedby many measuresof an Edwards re-
printof a Breitkopf& Hairtelpublication of Mozart's TwelveDuets for Two
WindInstruments,K. 487, whichhas servedas the sole source for Prerau'sex-

3Some of these algorithmsare presentedin my 'A systemfor the automatic reduc-


tion of musical scores', Papers Presented at the Seminar in MathematicalLinguistics,
vol. 6 (1960), on depositat WidenerLibrary,HarvardUniversity.
4Both dissertationswere supervisedat least in partby ProfessorMurrayEden.
5I call a written-musical compositionmonolynearif its sounded-musicalequivalentis
monophonic.In monolynearmusicthereare no chordsor hiatuses,so that in generalthe
sum of the note-durationsof each note and rest in a measureof monolynearmusic is
equal to the value of the time-signature
affectingthatmeasure.
6Prerau'sacronymfor 'sharp or flat or natural'(but not the double-sharp,whichhas
beenexcludedfromthedialect,and possiblynotthedouble-flat
or a sharpor flatpre-
cededimmediately
bya natural).

S252

This content downloaded from 128.192.114.19 on Sun, 31 May 2015 16:48:36 UTC
All use subject to JSTOR Terms and Conditions
COLLOQUY AND REVIEW

periments.The recognitionrules that Prerau adopts (for instance,to distin-


guish amongst 'sofons': 'The part of the componentcontainingthe top point
is the rightmost2/3 fora sharp,the leftmost1/3 fora flator natural;the part
of the component containingthe bottom point is the leftmost1/2 fora flat,
the rightmost1/2 fora natural' (p. 210)) thusare derivedfrommeasurements
of the same music Prerau's programproceeds to recognize; nonetheless,the
reportedexperimentalresultthat all dialectallypermittedsymbolsin twenty
or so measuresfromthe Edwardsreprintof K. 487 were recognizedcorrectly
is an impressiveachievement.His program,called DO-RE-MI, is writtenfor
the IBM 7094 computer,and produces output in the Ford-Columbiarepresen-
tation of CCMN,7 fromwhich,at least potentially,anotherprogramcan com-
pute the representedpitches. DO-RE-MI is said to be constructedso that
insertionof furtheralgorithmsto recognize portions of CCMN not included
in Prerau's subsystemcan be accomplished withrelativeease: however,these
algorithmshave yet to be conceived.
The question 'How simple is it to extend a computerprogramthat rec-
ognizes music issued by a certain firmat a certaintime so that the program
also recognizesmusic engraved(or printed)by otherfirmsor at othertimes?'
is closely related to the question 'How many 'founts' of music-engravers'
punches have been widely employed?': unfortunately,neitherquestion has
yet been answeredby any systematicinvestigation.But I understand,froma
valuable conversationwithMr. H. Edmund Poole who is much concerned with
certain historicalaspects of music printing,that much German, U. K., and
U. S. music engravingof the late 19th and early20th centuriesutilizedpunches
and other tools produced by Messrs. C. G. R5der of Leipzig. Any resultant
standardizationof manufacturingstyle can be only of benefitto the designer
of the ultimate OCR systemforrecognizingall printedCCMN, who will have
to plan not only fora varietyof different'founts'but also formusic symbols
of differing size, forabsence of expected spacingbetween symbols,forvarying
qualities of ink and paper, etc. The usually cited literatureon music printing-
e. g., WilliamGamble, Music Engravingand Printing(London, Pitman,1923)
or Karl Hader, Aus der Werkstatteines Notenstechers(Vienna, Waldheim-
Eberle, 1948)--will providehim only with littlehelp.8
Perhaps the greatestaccomplishmentof the authorsis that, as a resultof
theirwork, the logic of a machine that 'reads' multipleparallel staffsbearing
polylynear9printedmusic in at least one 'fount' and size can be seen to be no
furtherthan another couple of M. I. T. dissertationsaway. Quite possibly
such dissertationsmay get completed beforemuch thoughtis directedtoward
decidingwhat wiselyto do withthe masses of musical data that an operational
7See StefanBauer-Mengelberg, 'The Ford-Columbia inputlanguage', printedon pp.
48ff.oftheAmerican Musicological mentioned
Societypublication in footnote2.
81ngeneralPruslinand Prerauappearmerelyto haveskimmed theliteratureon mu-
sicaltopics.Whilstthishas notretardedtheirprogress,
careful readerswillfindpassages
inbothdissertationswhereforwantofmusicalknowledge fullinterdisciplinarity
wasnot
achieved.
9Thewritten-musical
analogueof'polyphonic'.

S253

This content downloaded from 128.192.114.19 on Sun, 31 May 2015 16:48:36 UTC
All use subject to JSTOR Terms and Conditions
PERSPECTIVES OF NEW MUSIC

OCR system could make available for computer processing. It is remarkable,


nonetheless, that (of all things) this technology may cause return of musicol-
ogists' attention to the core concepts of their field which constitute musical
theory.
Murray Eden has included a summary of Dr. Pruslin's work in an article.10
Dr. Prerau presented salient parts of his work in a report to the 1971 Fall
Joint Computer Conference.11 This report was judged the 'best paper' sub-
mitted and-by way of recognition-he was awarded a plaque.

-Michael Kassler

1Murray Eden, 'Other pattern-recognition problems and some generalizations'in


Paul A. Kolers and MurrayEden (eds.), RecognizingPatterns(Cambridge,Mass., M. I. T.
Press,? 1968), pp. 196ff.
11David S. Prerau,'Computerpatternrecognitionof printedmusic',Proc. Fall Joint
ComputerConference,vol. 39(1971) (Montvale,N. J., AFIPS Press),pp. 153ff.

S254

This content downloaded from 128.192.114.19 on Sun, 31 May 2015 16:48:36 UTC
All use subject to JSTOR Terms and Conditions

You might also like