Page  438 ï~~A Knowledge Based System for Recognition and Transcription of XVIth Century Guitar Tablatures D. Derrien-Nden Telecom Bretagne BP 832, 29285 Brest Cedex, France derrien@enstb. enst-bretagne, fr Abstact A frame-based and rule-based expert system for the automatic transcription of french sixteenth century guitar tablatures into standard notation has been developed. A related part of the project concerns the automatic recognition of printed tablature from scanned materials. 1. Introduction The task of old tablature transcription, like that of translating Chinese to English poetry, may appear unnecessary as it involves a considerable amount of interpretation and therefore some loss of information. However, just as everyone does not read Chinese with ease, the number of those who can read and play directly from old tablatures is rather small. In addition, Renaissance four-course guitars are rare nowadays. Furthermore, tablature is not a musical notation and therefore it does not provide an adequate representation for studying or analysing musical structures and rules of composition of this period. The automatisation of transcription is easily understood seeing the number of collections to study and the tedious nature of that task: generally, the tuning of the instrument was not indicated on the tablature and musicologists have to do as many transcriptions as hypothesis to select, at last, the most consistent tune. In this paper, we present an expert system which can help musicologists in the transcription of French guitar tablatures into standard notation. Rules implemented in the expert system for dance transcription were defined by Ms Helene Charnassd, musicologist at the ERAlTO research group at the Centre National de Recherche Scientifique (Paris). In the first part, we describe the tablature notation and the musicological problem of transcription. Then, we explain the structure and implementation of the knowledge base and the production rules. Last, we talk about tablature optical recognition which avoid the tedious task of material acquisition. 2. The transcription problem The four course-guitar was a standard instrument of the sixteenth century. It had four pair of strings and was a small instrument with a short string length. The tablature shown on figure 1 is an example of the ordinary French system also used in England during this period (german and italian notation were also used). The four lines on the tablature represent the four courses, and the letters, placed on top of any of the lines, indicate the frets to be fingered on that course (a = open string, b = first fret, and so on). Rhythmic symbols are placed up to several letters in vertical alignment which make up a chord; it remains valid for all succeeding letters until replaced by another rhythmic sign. Â~ " - _ -. b Rcmicr Ianlcc Bourgongnc. figure J A French guitar tablature Transcription reveals serious cognitive and technical problems. Each stringed instrument (lute, guitar, viol, harp) has a specific notation which varies according to the country and the period. Consequently, tablature understanding requires the competence of different specialists (musicologists, palaeographers, organologists...) capable to decipher this complex notation. Futhermore, implicit factors of rendering, used by Renaissance instrumentalists have been perceived during notation analysis. In fact, a great amount of logic and oral tradition takes place in its comprehension: some notes, obvious for the performer of that period were omitted. Then, the goal of transcription is to rediscover the internal logic of the notation that was well known by 16th century musicians and that has partly disappeared with the loss of tablature usage. 438 ICMC Proceedings 1993

Page  439 ï~~There are six main tasks in tablature transcription: " data acquisition " simple transposition of characters into notes " selection of the alteration of the notes " search for the polyphonic structure " compute the duration of the notes " display notes on two staves Each of these tasks will be detailed in section four, except data acquisition which will be introduced latter. 3. The Knowledge Based System In musical transcription as in most cognitive problems, the characteristic properties of the objects we are manipulating are not always known a priori. So, we have to create, then kill or bring up to date various objects; to add or suppress different properties; to structure those objects in hierarchies by means of "isa" or "part-of" relationships. Therefore, knowledge modelling is better structured using a frame based system rather than a Prolog or Lisp language. So, we chose the frame-based system Y3 [Ducournau 91] to design our evolutive knowledge base on tablatures. YAFOOL, the kernel of Y3 is implemented on the Le-Lisp system. This hybrid system also contains the rule based inference engine PRYSM, the pattern-matching tools FYLTRE and the graphical user interface YAFEN. That interface allows the visualisation of application behavioural evolution and therefore makes prototyping easier. The display of tablatures and scores was also realized with this interface (see figures 2 and 4). Both tablature and score layout structures are represented in hierarchies using "part-of" links between the various entities appearing in a musical work. Each entity is associated with a Yafool frame (group of lines or staves, bars, verticals, letters or sounds...). Each musical entity has also logical links between its logical neighbours (independently of layout cuts such as change of page or stave). For example, the musical-symbol frame is described by the properties next-symbol, previoussymbol, part-of (the bar which contains this symbol) and string (the string to be played). It is speialized in two specific frames sound for scores and fret for tablatures. Sound frame inherit all musical-symbol properties and is speialized with the properties height, octave, duration and note. It is itself speialized into twelve pitches with predefined note slot (for example the frame sound2 has two possible values for its slot note: C# or Db). The fret frame is described by the properties chord (a list of couple <sound, octave> indicating the sound associated to each string for this fret. This slot is filled when the tune is given). Another example of modelling is given by the slot rhythm in the vertical frame. If its value is not available a deamon if-needed enable its inheritance from its previous neighbour. This slot is typed: its value must be a specialisation of the frame rythmic-value. 4. Rules for Transcription The production system PRYSM integrated in Y3 is based on the following caracteristics: " one can define a collection of condition-action rules called productions " the knowledge base is independant of the rules: computations are carried out by successive revisions of this knowledge base. " the interpreter works in cycles: each cycle decide which rule(s) to execute, and then execute them According to the rules defined by Ms Charnassd, productions have been added to the knowledge base to perform the transcription. 4. 1 Simple Transposition The first part of the automatic transcription deals with the direct transposition of alphanumeric symbols into sounds described by their height and octave, on two staves. This task is implemented in a simply way, once the tuning of the guitar is chosen (the chord slot in the afret frame is then updated). A deamon attached to this slot enable the automatic updating in the other frets. 4.2 Identification of altered notes French guitar tablatures does not contain accidentals (flat or sharp). Consequently, during the translation we have to determine the correct note (is it an A# or a Bb?), once for the entire work. This approach is based on the musical context: to determine the accidental note a rule studies the value of the surrounding notes in this chord and in its neighbours. For example: Rule A: If a vertical contains an altered note (C# or Db) and also a note among (F Ab Bb) then the note is a Db 4.3 Retrieving the polyphonic structure This study is based on the concept of <<chord saturation > when all voices should participate. Indeed, this music was originally intended for a vocal ensemble, so in saturated chords the number of notes indicates the number of voices. Consequently, rules based on the ICMC Proceedings 1993 439

Page  440 ï~~counterpoint are defined to guess the continuity of melodic lines between two saturated chords. Only the upper melodic line was studied actually. 4.4 Rhythmic symbol interpretation Rhythmic symbols on the top of some verticals indicate only the duration of the shortest sound in that vertical, so other sounds can have a longer duration. This calculation is done for each melodic line. In a voice a note is sustained until another note appeared or according to physical properties of vibrating strings and harmonic conflicts. Therefore, several rules have been implemented to compute note duration. To display the result, each sound duration is decomposed into admissible rhythmic structures and presence of bars is taken into account to establish ligature. Final result is shown on figure 2. Rules are only implemented to retrieve the upper melodic line, so duration of notes belonging to another voice are not calculated and are displayed as minim. tJ~~~~~ ML O u 0 figure 2.Result of the transcription of fig. 1 tablature 5. Tablature Optical Reading In the first version of our system, the tablature handling was performed by means of a text editor to encode each musical symbol. The structure of the work was described by nested lists. Then, we considered optical recognition to avoid this tedious task to musicologists.Tablature recognition is performed using a top-down analysis. First the physical structure of each page of a collection is recovered and then each musical symbol is recognized. The first step of structure analysis concernsthe dermopialeontion oupso thizota ledos inseach mscndopget.Thisu ecito is performedbysuinth pofilchae of rzal ojection reofee all black piels. Pea on'tht stpofletindicatesnthesapproximate lin location. Then for each group of lines, bars are detectedl and removed according to the vertical projection of their black pixels. The second step deals with lines elimination: as lines are slightly bent, the elimination is done by contour following rather than blanking out its position on the entire page width. A little window is placed upon each line to scan it and delete black pixels contained in its thickness. When intersecting a letter the window is shifted to the right. Â~ -V c -a f f h f "D4 aC 4 C. figure 3.Result of fig. 1 lines elimination Once line elimination is achieved we extract the connected components of the tablature. Location of chords is computed by vertical projection of these symbols. Then, each connected component is recognized using a decision tree, and finally the layout structure is build and passed to the knowledge base. - -., - I ", " - --,-b7... figure 4. Result of fig.1 optical recognition 6. Conclusion In this paper we described some rules to automatize the simplest (and most tedious) parts of the transcription problem. However, it is clear that this task could not be resolved entirely in a logical way with production rules. Musical composition is part of artistic world and not of mathematical or logic world.Consequently, such a system should be considered as a working tool for musicologists and not as an autonomous software. Rules were implemented for dance transcription [Le Roy 1989], and for French guitar notation system. However the system is easily extensible to other notation. Other rules should be added to build all melodic lines in the same way. 7. References [Charnass 901 Chamass6 H., Stepien B.,, Automatic transcription of german lute tablatures Â~ an artificial intelligence application >, Mardsen & Pople (eds), 1990. [Ducournau 91] Ducoumau Roland, ((Y3. YAFOOL, le langage?i objetsx., Manuel de drfrence (version 3.64), Sema-Group, 1991. [Le Roy 89] A. Le Roy, R. Ballard, (<Cinq livres de guitarre, 155 1-1555 >, Edition Chanterelle, Monaco,1989. 440 ICMC Proceedings 1993