Back to Tim Crawford's research
There is nothing new in observing a relationship between an aptitude for music and one for mathematics: it was formally recognised in the mediaeval curriculum, and is a subject of some fascination for cognitive psychologists and their like in our own age. Something similar I have noticed over twenty years spent playing and investigating lute music is the surprising extent to which computing and lute-playing seem to be compatible activities.
In a recent poll of its membership, the Lute Society of America found that no fewer than 76% of members considered themselves to be computer-literate. While this is obviously a very subjective self-evaluation, the high figure does, I think, reveal a degree of genuine statistical evidence for my own observations: viz., that there is a predeliction for the use of computers among a group (lute players) who might reasonably be assumed to be out-and-out Luddites in their attitude to technological aids of this sort. After all, these are the people who comb old sources for information about the right kind of sheep from which to extract gut for their `authentic' strings; who refuse to countenance the performance of a piece on a historically inappropriate type of lute (out of the twenty-or-so varieties), let alone on the dreaded guitar; and -- to outsiders perhaps most mysterious of all -- who insist on playing and even performing from facsimiles of original sources, even when these are full of misprints or scribal errors. Yet many amateur lutenists of my acquaintance are highly-skilled computer users and/or programmers: they range from a consultant called in to supervise a major installation for the Hungarian government (recently, I am glad to report) and a Ford designer who installed the first driver-oriented computer system in a UK production car, to the developer of one of the most successful and versatile music-notation programs available for the Macintosh computer.
On the other hand, the musicologists who work in this area are, with some honourable exceptions, rather resistant to the new technology. In some cases, there is some justification: many of my colleagues have pointed out that the human brain's ability to collate information in several directions at once combined with its wonderfully `selective' memory capabilities has enormous advantages over current computing methods in this area of study. But only one researcher in this camp to my knowledge has that essential personal gift that only an as-yet-undeveloped computer of the future could substitute for, a `photographic memory' for music that really is reliable. This particular scholar has the further advantage that he is a highly accomplished lute player himself.
While I admit to a degree of personal bias -- `Of course lute scholars should be lute players!' -- it is clear that a lutenist has an enormous advantage over a non-player, in that the notation of lute music is as clear to the lutenist as that for keyboard, let us say, is to other musicologists. Tablature -- the lute's special notation system -- has obstructed the real progress of scholarship in the lute field ever since the very earliest days. Only an experienced lute player can look at a piece written in tablature and `hear' the music as I am sure most of you can `hear' a printed score in conventional notation.
I suspect that the reason for the rather surprising ease which lutenists feel with computers is precisely concerned with the nature of this tablature notation, which is rather like an algorithmic `program' for the performance of a piece of music. Its tabular nature, its use of normal text characters rather than specially-shaped `glyphs' for the notation of basic notes, and its representation of the fingerboard of the lute and the placement of the left-hand fingers have all ensured that the principle endured throughout the lute's history (even to the present day in the case of certain guitar notations which seem never quite to have died out), and that it is almost universally preferred by lutenists even when performing in concert today. It also leaves open to the player himself or herself the widest possible range of interpretational possibilities: even to the extent that in some cases -- not just in the infamous unmeasured preludes -- the relative durations of the notes are not set exactly by the notation. These are things that are impossible to notate explicitly in conventional notation, and perhaps are best left implicit in any case.
Accurate transcription is not only tedious, it is also full of pitfalls for the unwary. Personally I further believe that it is a chimera: an unattainable goal. No two scholars or players will agree about every aspect of the methodology of transcription, nor about the precise interpretation of every symbol in any given tablature.
I am of course aware of the research going on into systems for the automatic transcription of lute and other tablatures, and can see that the ingenuity of the procedures is impressive, but it seems to me that these are very valuable exercises in the application of expert systems and other advanced technologies to the interpretation of a given input data-set according to sets of rules that apply only to a limited repertory (or subset of the possible range of input data-sets). Furthermore, the rules themselves are of necessity designed according to `received' academic interpretations of contemporary theoretical procedures. [As a trivial illustration, if a scholar investigating the 16th-century lute fantasia believes that parallel 5ths or octaves would have been anathema to a lutenist's theorist contemporaries and that therefore they should be avoided in his transcriptions (manual or automatic), his transcribing technique will most probably be wholly inappropriate to the dance music of the time. Furthermore, the very notion of contrapuntal integrity of voices in lute music is one that can only very rarely be sustained; there is also a vast `grey area' of very fine and important lute music whose style I would argue is based on a living tradition of improvisation and could never be forced into the mould of academic rules in any case. But this is an argument that must be pursued on another occasion.]
For this reason, while wishing in no sense to detract from the excellent work that has been done in automatic transcription and hoping that those involved will not become disillusioned by the complexity of their task, I cannot but remain sceptical that this theoretical-rule-based approach will ever yield a general tool for musicologists to use for treating lute music. It must be recognised that the repertory of primary interest for any one researcher will strongly `colour' any rule-based method that researcher may adopt: rules of counterpoint, motivic elaboration, harmonic rhythm and metrical structure assume different degrees of priority in the study of the diverse repertories of the dance, the fantasia and intabulated vocal music in the 16th century, while the basis on which they are assessed must be different when looking at analogous repertories in the 18th century.
In the necessarily brief survey of what I am myself attempting to achieve that follows, I shall be concentrating primarily on one important area where, despite numerous expressions of good intent over the last 85 years, lute music research lags well behind the achievements of the mainstream: the cataloguing of sources. The systems that dominate this field, MARC and RISM, were conceived without much -- if any -- consideration for the special problems of tablature sources. I hope that the principles involved will prove useful in other sorts of lute-music research, style-analysis in particular, and may even suggest an avenue of possible progress for dealing with other repertories where formal polyphony is not a high priority and single strands of melody are hard to extract.
While there is clearly an interesting topic of pure research here, my own participation was inspired by my more-or-less simultaneous involvement in three significant developments, the beginnings of the CDSMT cataloguing project, my appointment as editor of the complete works of the great 18th-century lutenist Silvius Leopold Weiss, and the start of my involvement in beta-testing the Macintosh music-notation program Nightingale under development by Donald Byrd.
The most directly relevant to the present discussion is the CDSMT (Catalogue Des Sources Manuscrits en Tablature), which is a determined effort to catalogue properly all sources of music in tablature for lute, guitar and related instruments.
Although the CDSMT has begun to bear fruit in the shape of some published inventories of lute music manuscripts in Swiss and French libraries, one of the original aims, to set up a centralised database of lute music incipits, has still to be realised. Having been involved in technical discussions at an early stage, it was obvious to me that currently available database management systems, while they could more than adequately deal with the standard bibliographical aspects of a catalogue, were not suited to dealing with musical incipits in any useful manner. Nor could lute tablature usefully be integrated into a scheme without some means of representing its appearance, and, if possible, its musical implications. For the reasons I have just outlined, I am certain that cataloguers should restrict themselves to recording the tablature in the source, and not actually transcribe it, even within strict guidelines. It had been decided at an early stage that the tablature music incipits should be recorded on work sheets, and that this data could be incorporated into the database when it could be accommodated. As far as I am aware, this principle is still in place.
My own previous research had involved developing a crude cataloguing system (in a Macintosh development environment called HyperCard) whereby French tablature could be entered into a database in an encoded form, and also be easily retrieved and its appearance reconstructed in `diplomatic' form without too much difficulty. Any tuning of the lute's playing strings (or those of the lyra-viol, for which the system is equally appropriate) could be accommodated. The tablature could be `played back' through a MIDI system, and edited to eliminate errors of data-entry and so on. Various data-fields could be extracted from the encoded tablature, including a simple listing of the highest-sounding notes of each `sound-event' or chord, which could be `normalised' to C major/a minor in the familiar way to offer a rudimentary index for searches and sorts of the database. The principle was found to work in a useable way for restricted repertories, but lacked the kind of flexibility and sophistication necessary for a full-scale project. Although some preliminary work was done on the design of a database system for a CDSMT cataloguing scheme, the lack of funding meant that this could not be carried far, and no further development of actual incipit-processing techniques could be achieved.
In order to discuss some of the special problems involved in the computer-based cataloguing of lute music, it will be necessary to give here a brief summary of the basic workings of lute tablature. I shall be dealing here only with French lute tablature since I regard it as axiomatic that there is nothing in German, Italian or Neapolitan tablature that cannot be represented in the French form.
Tablature represents the six main playing strings of the lute as six `staff-lines' onto which letters are placed signifying fret-positions of the left hand fingertips. An `a' on a line means the string is played open, `b' signifies the first fret, and so on. The relative duration of each vertically aligned sound-event, or chord, is indicated by conventional signs based on normal musical notation placed above the `staff'.
There is a major difference in principle between tablature and conventional notation. Lute tablature is best described as a prescriptive system of notation as opposed to the descriptive system familiar to most musicians. [In this it resembles the MIDI protocol, which -- at its simplest -- merely tells a device when, on what channel and how loud to sound a note.] It indicates to the player where to put his fingers and when (in relation to other sound events) to strike the strings. It does NOT seek to represent the music per se in any theoretical or physical sense, unlike conventional notation, which shows high notes at the top of a staff, low notes at the bottom. In durational terms it only indicates the onset of each note, leaving decisions about the end-time of each note (in other words the duration of each chord-member) entirely to the player or to the physical demands of the instrument (for example, a new note on a string necessarily implies the earlier termination of the previous note on that string.)
There's no doubt that this prescriptive character accounts for much of tablature's appeal to partially- or un-trained musicians (computer buffs among them!) who when learning the lute need tackle only the most rudimentary elements of rhythmic notation to produce music -- there's certainly more than enough else to learn! -- and can thus sightread very early in the learning curve. [This is also borne out by the high number of lute MSS written out for amateurs, which contain explanations of rhythm as their only explicit didactic content. The preface repeated in each of Petrucci's solo lute publications is of the same nature.] In spite of the apparent disadvantages of the tablature system (it cannot, for example, accommodate triplet/duplet simultaneities: ie playing 2 against 3) it is quite adequate for performers, who positively relish its vagueness in certain aspects, a vagueness and a relish which I am sure relate to the un-notatable element of improvisation that is never far from the surface in the best lute playing.
For a thorough and systematic bibliographical project like the CDSMT database, it goes without saying that as much as possible of the appearance of the original source should be recorded. But a truly diplomatic method of recording lute tablature incipits (ie one that shows enough of the original appearance to transmit all its possible musical implications) is harder to achieve than it may seem at first sight, for there are many other things present in most tablatures than the basic elements I have just described. These include ornament signs, text (whether actual lyrics to be sung, titles or comments) and various other graphical devices, from lines and slurs to petite reprise markings and even pointing fingers in the margin that may or may not have musical significance. And a comment like `NB belissima' hand-written against a piece by the owner of a 16th-century print may have an important message to tell some researcher. For this reason, the draft TabCode description that I can supply to any interested person, may seem unecessarily complicated for such a simple notational system. But it may actually be too simple; I need more comments on this.
At present TabCode consists of a simple code (`Minimum TabCode') which can be elaborated to produce a complete encoding of all the graphical elements (`Full TabCode'). The extra `diplomatic' data for Full TabCode is either placed in normal round parentheses, or comprises special mnemonic symbols for fingerings and the like, so it can be rapidly eliminated for data-processing, which is probably more efficiently performed on the Minimum TabCode.
The principle for producing the Minimum TabCode is very simple: after encoding the initial time signature using a signifier `M' for metre and an easily recognised form like `C' for common time or `C/' for cut time, each musical event (or Tabword) is encoded vertically downward, beginning with a mnemonic code for the rhythm-sign (ALWAYS a capital letter and based, as usual, on the US notename, `W' for whole, `H' for half, and so on) concatenated with the letters and string numbers for each of the notes of the chord. So `Wa1b3c4' represents a whole-note chord comprising an open first string, the third string stopped at the first fret (`b') and the fourth stopped at the second fret (`c'). Barlines are represented by the vertical bar symbol `|'. Tabwords and other data-components are delimited by space-characters.
TabCode provides an economical and, I hope, unambiguous means of encoding French lute tablature. For the present, Minimum TabCode seems to work well, while there are some problems and deficiencies in Full TabCode and some features yet to be implemented, such as a means of specifying the orginal tablature system of the source, which may be German or Italian tablature, in such a way that the original can be reconstructed for display or printing. I have not begun to think about ways of representing the various guitar tablature systems, including the chord-shape system using capital letters known as alfabetto, and which is found in a large number of manuscript sources. There are also some devices I have not included which only occur in a very few MSS, such as the notation for the `split mezzana' in the Capirola lute Book, where the two strings of a double-course need to be played separately. Incidentally, TabCode is just as useful for the encoding of music written in tablature for the lyra-viol or viola da gamba, the baryton, or even the violin.
In order to reduce the tedium of tablature encoding, which is no worse but no better than any other musical encoding, I am also working on a rudimentary Tablature Processor which allows the user to type the tablature symbols directly onto the staff using a specially designed font. The music can be played back, and saved in a TabCode ASCII file from which the tablature can be printed or imported to a database field for bibliographical purposes.
Since tablature is merely a visual prescription of a `program' of actions to produce a desired piece of music, it presents fundamental differences of treatment and philosophy from a theory-based descriptive system. Apart from the matter of individual polyphonic note durations, which is the most obvious difference, there are some more subtle incompatibilities which must be faced.
For the rest of my discussion I shall mainly be considering Baroque lute music in which we can assume that `modern' major and minor keys are fully established. While this is a gross simplification, even for some late 17th-century lute music, I feel sure that similar approaches for lute music using `modal' techniques could be fruitful.
A vertically-segmented notation system like tablature could reasonably be expected to provide a useful description of the harmony of a piece; indeed, this is so, if harmony is simply considered as the progressively changing pitch-class content of the sounded chords. But this is a quite inadequate view of traditional tonal harmony, since it ignores the vital horizontal concepts of voice-leading, cadences and so on.
As with MIDI there is no such thing as enharmonic spelling in tablature (except in so far as the sounding pitch of open bass-strings is sometimes named or implicitly calculable: in other words, one knows that the open 11th course in a piece for Baroque lute in D major will be tuned to C#, not D flat). Most frequently the actual `key' of a piece is not given, and is actually a relative term in any case, since -- for solo performance at least -- lutes could be, and were, tuned over a wide range of pitches.
The interval tuning of the lute's six basic playing strings varied considerably over the lute's history: the familiar Renaissance tuning in 4ths with a 3rd between the third and fourth strings was in use throughout, though in the few remaining decades from about 1740, it was only used in Italy. A profusion of `scordatura' tunings arose in the second and third decades of the 17th century, one of which, based on a minor chord with the third at the top, became a virtual standard later in the century. In around 1800, at the very end of the lute's life, a major-chord tuning based on that of the so-called English guitar was used, although there are no tablature sources for the hybrid instrument that was used at that time.
Sometimes, however, the tunings are not specified, and in some sources pieces are jumbled up in such a way that pieces in very different tunings are juxtaposed - this is the case for two MSS recently coming under close scrutiny: one from the Cortot collection now at the Newberry library, in Chicago; the other at the Academy of Sciences in St Petersburg. Any computer-based system that attempts to use musical data derived from tablature had better get the tuning right in every case or the results will be meaningless.
There may be a possibility of using the computer to identify the tuning of a given tablature sample by examining the occurrence of frequently-used chord-patterns. But without such an aid, which can only be a long-term objective, the provision of MIDI playback would seem to be the best way of testing likely tunings, since changing the default tuning is a fairly trivial matter.
Once the tuning has been established to the satisfaction of the user (and, let us not forget, a concomitant element of subjectivity introduced into the process) it remains to decide whether the piece is major or minor, which possibly could be done, since lute music in general modulates very little, by a statistical analysis of scale-degree occurrences related to the final chord of the piece. This could work well for the majority of baroque lute pieces that end with a root-position chord, or a bass note alone, but there are plenty of anomalous cases, and plenty of `corrupt' sources that could mess things up. It would also fail for pieces lacking their final chord. In these cases, the researcher might have to swallow his technocratic pride and use his ears. However, a reasonably pragmatic solution would be to provide a manual override to this rather cumbersome procedure, since the key is usually pretty clear in the common tunings.
The main object of incorporating a coding system like TabCode into a bibliographical database like CDSMT's or RISM's would be to search for `concordances'. There are basically three types of correspondence between sources that we need to consider:
i) Strict Concordances: other copies of the same music, identical in all significant respects or with only minor deviations in detail. (e.g. The manuscript version of a Presto in Bb by Weiss, and the version printed in Telemann's Der Getreue Musikmeister differ only in one or two notes and a few slurs)
ii) Cognates: other versions of the same music, differing in substance but ultimately derived from the same composition (A very few of the 20 or so manuscript copies of Dowland's Lachrimae Pavan present concordances of one or two apparently authentic versions; most are arrangements of the cognate type which maintain Dowland's outline while significantly altering inner parts or providing different divisions.)
iii) `Other settings': these would include arrangements of the same music for other instruments, as well as complete re-settings. (Examples might include different composers' treatments of popular tunes or bass-patterns)
These three degrees of similarity can only be expressed in fairly vague and subjective terms. An implementation of automatic searches for them will of necessity be an extremely complex computing task. For the present, it will probably be most fruitful simply to look for common features between incipits and rank them according to a scoring system such that an absolute identity scores 100, with divergences suffering weighted penalties until a cutoff value is reached that counts as `no match' for the purposes of the scan.
Of course an intelligently designed relational database will already have eliminated a vast proportion of the irrelevant material before this computationally expensive searching begins: I am most unlikely, for example, to want to search for music by Weiss in a 16th-century MS. (However, one lute piece popular around 1600 consistently appears in manuscripts for more than the next century: its last appearance seems to date from around 1750!)
It would be necessary to consider two basic types of concordance search:
i) those conducted within the database of lute music in order to establish internal concordances
ii) searches done in order to establish concordances with other repertories, eg for keyboard or ensemble music already stored within the RISM database.
At first sight these may seem to involve entirely different procedures -- in one case lute music is being compared with lute music, in the other the comparison is across different media -- but it is unlikely that even for the first type of search, involving lute music only, a simple search for identities in the TabCode would be at all useful, so similar approaches may be needed in each case.
In the case of an `internal' search, generally speaking, it is likely that any variant in the original tablature will produce a change in the TabCode of a similar order. For example, a note changed to another fret position on a lower string will involve changes to no more than the same two values, string number and fret letter. So some sort of crude measure of dissimilarity between sources could be achieved by counting differences in the TabCode -- a task for which a simple text comparison program would suffice. But it would be better to rank variants by assessing the likelihood, for example, that one source simply has the right note written on the wrong string or that the wrong rhythm sign has been copied. So that variants where only a string number or a rhythm character is changed might incur a smaller penalty in ranking than the case where a tablature letter is different. (This type of `weighting' has a parallel in the `distance' weighting between compared melodies suggested in a recent study by Sankoff and Mongeau in Computing and the Humanities, where the degree of dissonance between substituted notes contributes to a score of dissimilarity.)
More seriously, it would be well-nigh impossible to detect transposed versions, for example, using TabCode in its raw text state. A more sophisticated approach is clearly called for.
For `external' searches it would seem to be more fruitful to regard them as a special case of the internal type; in other words, to regard a monophonic incipit as a kind of tablature in which each Tabword, or chord, contains only one note. Polyphonic external search-objects could be `intabulated' in an internal structure analogous to the vertical segmentation of tablature and then compared in the same way as in the internal concordance search.
A related question is that of the integration of lute music incipits into the RISM scheme of cataloguing. The present RISM system only allows for single-line monophonic incipits to be stored and compared, and has already proved its worth in searching for concordances in repertories such as anonymous 18th-century symphonies. But, since it does not even allow chords, there is no satisfactory way to enter lute music.
The CDSMT project has as one of its goals the aim of integrating its catalogue with RISM's, so that the extent of cross-concordances (if I may so describe them) between lute, vocal, ensemble and other instrumental music can be assessed. It is hard to see how this can be achieved without some radical change in RISM practice. Hopefully this can be achieved in such a way that the enormous number of incipits that have already been entered are still of value, yet `informally polyphonic' music such as that for keyboard, lute, viola da gamba or even unaccompanied violin could be usefully accommodated. I would suggest that tablatures should be treated specially in the sort of way I have hinted at. German -- and indeed Spanish -- keyboard tablatures, which I have not researched in any depth, might also be treated in a parallel way, though I acknowledge that they may produce other problems.
There are futher problems that might be anticipated in studying Baroque lute music. These derive from the interaction of at least three special features of the repertory: i) its notation, ii) the nature of the sources and the subsequent transmission patterns of the music, and iii) its performance traditions.
Many, indeed most, Baroque lute manuscripts served a complex of artistic, social and didactic purposes. Frequently compiled by, or under the supervision of, expert professional teachers, they cannot be treated simply as simple objective records of the music - those ideal primary sources we all hope to find - owing to the performance-related prescriptive nature of the notation. The lute had a persistent tradition of extemporisation, and it is clear that composers for the instrument felt a conflict between this tradition and their natural desire to have their music transmitted in copies which reflected their artistic intentions in the composition. But extemporisation, or at least its effects, tended to win out in the sense that individual sources of a piece, while containing the same actual notes, can show a wide variety in the way certain performance features are notated. The most obvious and well-known example of this is the range of differing ornament signs used by different scribes. Often one ornament is substituted for another, with a somewhat different effect, though only on the level of detail.
Another widespread feature of Baroque lute music, one that was often imitated in keyboard transcriptions, was the traditional habit of breaking chords rhythmically; a device that was known at the time as the `style luthée'. In an inexpert keyboard transcription this can give the impression of a highly broken melodic line with little internal logic; the modern designation `style brisée' reflects this character without fully appreciating the fact that keyboard notation cannot give this kind of performance indication without a proliferation of ties and syncopations which thoroughly confuse the structural form of the music. The intended effect was more probably nearer to a controlled rubato in the various evanescent polyphonic voices than the deliberate defocussing and romantic blurring that has been suggested by a few modern commentators, and indeed achieved by some modern players. In tablature, this spreading is either not indicated at all (a situation that may represent in some cases a kind of basic form of a piece, but one which may never actually survive in the sources), is indicated by special signs (slanting `separée' lines between the chord members) or is shown by writing out the broken chords as they should be played. If we accept, as I think we must, that these variants between sources should not disqualify two versions from being considered true concordances, there is a real difficulty here for an automatic concordance recognition system. It means that if an item in the search-object is not found precisely where it `should' be in the target, it is necessary to consider its possible temporal displacement to a nearby position according to guidelines which will be difficult to specify unambiguously. Some kind of rhythmic/harmonic segmentation on a chord-to-chord level will be necessary to achieve a useful implementation of this kind of strategy. But once this has been achieved, it is possible that similar techiniques could be put to good use in other types of melodic recognition where similar displacements can be expected.
I have identified several problems which have arisen from my study of the general question of automated melodic recognition and comparison in lute music:
It is obvious that none of the solutions to these problems could be described as trivial, and that the programming techniques required will have to be highly sophisticated. It is clear, too, that it would be useful to build on the experience of those working in other disciplines. Computer scientists have suggested to me that some techniques associated with artificial intelligence could be extremely helpful. The notion of `fuzzy' matches is well known in that field, and is directly applicable here; list processing and the potent combination of neural-net theory and parallel processing could also be useful. An especially promising area is the very advanced work that has been done on DNA molecular pattern matching, which has many parallels with the requirements of this research.
These problems will not be solved by any one person, nor is it likely that they will be solved in the very near future, judging by the past history of our subject. But they must be faced, and should offer interesting challenges to computer scientists and musicologists alike, but also -- and more importantly -- add to the understanding of an important and under-appreciated repertory, what Thomas Mace in 1676 described as THE BEST MUSICK IN THE WORLD.