The identification and modelling of a percussion ‘language’

James Kippen & Bernard Bel

Computers and the Humanities (1989), 23, 3: 119-214

Abstract

In exper­i­men­tal research into per­cus­sion ‘lan­guages', an inter­ac­tive com­put­er sys­tem, the Bol Processor, has been devel­oped by the authors to analyse the per­for­mances of expert musi­cians and gen­er­ate its own musi­cal items that were assessed for qual­i­ty and accu­ra­cy by the infor­mants. The prob­lem of trans­fer­ring knowl­edge from a human expert to a machine in this con­text is the focus of this paper. A pro­to­typ­i­cal gram­mat­i­cal infer­encer named QAVAID (Question Answer Validated Analytical Inference Device, an acronym also mean­ing ‘gram­mar' in Arabic/Urdu) is described and its oper­a­tion in a real exper­i­men­tal sit­u­a­tion is demon­strat­ed. The paper con­cludes on the nature of the knowl­edge acquired and the scope and lim­i­ta­tions of a cognitive-computational approach to music.

Excerpts of an AI review of this paper (Academia, June 2025)

Summary

This paper explores a nov­el approach to mod­el­ing North Indian tabla drum­ming as a “per­cus­sion lan­guage” by apply­ing for­mal lan­guage the­o­ry, machine learn­ing, and inter­ac­tive generative/analytic com­put­er meth­ods. The authors dis­cuss two sys­tems— Bol Processor and QAVAID — that each plays a dis­tinct role in ana­lyz­ing and gen­er­at­ing rhyth­mic pat­terns (termed “sen­tences”) under the guid­ance of expert infor­mants. They exam­ine how knowl­edge is incre­men­tal­ly acquired and for­mal­ized as a gram­mar, how alter­na­tive seg­men­ta­tions can be eval­u­at­ed, and how prob­a­bilis­tic mod­el­ing may be employed to gen­er­ate orig­i­nal musi­cal sen­tences for expert eval­u­a­tion. The work’s eth­no­mu­si­co­log­i­cal per­spec­tive unites com­pu­ta­tion­al for­mal­iza­tion with the real-world prac­tice of tabla impro­vi­sa­tion and teach­ing, rais­ing broad­er ques­tions about the nature of knowl­edge trans­fer between human expert, machine learn­er, and cul­tur­al context.

Contribution and Strengths

Interdisciplinary Framework

The paper posi­tions itself at the inter­sec­tion of musi­col­o­gy, cog­ni­tive sci­ence, com­pu­ta­tion­al lin­guis­tics, and ethnog­ra­phy. This breadth under­scores the com­plex­i­ty of “music as lan­guage” and effec­tive­ly high­lights the idea that music may be for­mal­ly scru­ti­nized with meth­ods akin to those in com­put­er science.

Formal Language Techniques

By ground­ing the analy­sis in the Chomskian hier­ar­chy (reg­u­lar and context-free gram­mars) and ref­er­enc­ing Gold’s con­cept of “iden­ti­fi­ca­tion in the lim­it,” the authors tie their eth­no­mu­si­co­log­i­cal obser­va­tions to well-established the­o­ret­i­cal under­pin­nings. These con­nec­tions help clar­i­fy why a sys­tem­at­ic, incre­men­tal approach to gram­mar infer­ence is suit­able for mod­el­ing the impro­vi­sa­tion­al com­po­nents of North Indian tabla drumming.

Attention to Vocabulary and Segmentation

The dis­cus­sion on how the sys­tem learns seg­men­ta­tion and defines “words” in the drum­ming lex­i­con is illu­mi­nat­ing. Though seg­ment­ing tabla phras­es is not anal­o­gous to seg­ment­ing words in spo­ken lan­guages, the authors show how incre­men­tal analy­sis can pro­pose, refine, or dis­card poten­tial lex­i­cal bound­aries in a prin­ci­pled manner.

Interactive and Incremental Learning

A sig­nif­i­cant fea­ture is the inter­ac­tive mod­el: the sys­tem gen­er­ates out­put strings that are val­i­dat­ed or reject­ed by the human infor­mant, there­by trig­ger­ing incre­men­tal adjust­ments to the gram­mar. This mim­ics student-teacher inter­ac­tions and demon­strates a strong attempt to reflect authen­tic learn­ing and teach­ing processes.

Probabilistic Aspect

Introducing sto­chas­tic­i­ty in syn­the­sis breaks from pure­ly deter­min­is­tic meth­ods. It points to a more real­is­tic reflec­tion of the ways in which live per­for­mance might involve cre­ative, non-deterministic choic­es, while main­tain­ing con­straints guid­ed by the learned grammar.

Methodological Observations

Data Representation

The authors clear­ly define the sym­bol inven­to­ry (bols like dha, ge, ti, etc.) and acknowl­edge the com­plex­i­ty of how these sym­bols relate to son­ic events. By lim­it­ing the approach to frequency-based seg­men­ta­tion and gram­mar infer­ence, the sys­tem oper­at­ing with­in a “text pre­sen­ta­tion pro­to­col” remains suit­ably rigorous.

User–System Dialogue

Illustrations of the QAVAID question–answer mech­a­nism high­light prac­ti­cal aspects of gram­mar con­struc­tion. This is valu­able for explain­ing how the sys­tem backs up, mod­i­fies rules, or infers new chunks based on par­tial dis­agree­ments from the expert and how it tests repeat­ed merges or seg­men­ta­tions for consistency.

Scalability Considerations

The exper­i­ments pre­sent­ed involve a lim­it­ed num­ber of exam­ples. The authors note com­pu­ta­tion­al con­straints and care­ful­ly frame how repeat­ed merges, lex­i­cal expan­sions, and neg­a­tive exam­ples (machine out­puts the user rejects) unfold in real­is­tic time on a micro­com­put­er. This trans­paren­cy about per­for­mance con­sid­er­a­tions is commendable.

Comparison to Existing Tools

While the authors ref­er­ence for­mal lan­guage the­o­ry, it could be help­ful to sit­u­ate the QAVAID approach more explic­it­ly along­side oth­er grammar-inference sys­tems (or music cog­ni­tion mod­els) in terms of effi­cien­cy and suc­cess rates. This might pro­vide addi­tion­al con­text about how QAVAID’s tight-fit method­ol­o­gy dif­fers from exist­ing machine-learning strate­gies in music.

Suggestions for Future Work

Integration of Connectionist Approaches

A deep­er inves­ti­ga­tion into how sub-symbolic learn­ing algo­rithms (e.g., neur­al net­works) might coex­ist or com­ple­ment a sym­bol­ic grammar-inference approach could shed light on whether deep­er hier­ar­chi­cal or pattern-based musi­cal struc­tures can be dis­cov­ered automatically.

Temporal and Metric Awareness

Incorporating real-time con­straints, includ­ing an explic­it mod­el of cycle bound­aries and tem­po vari­a­tions, might enable QAVAID or a suc­ces­sor sys­tem to han­dle per­for­mances that devi­ate sub­tly from rig­or­ous­ly mea­sured durations.

Generative Evaluation

Extending the sys­tem to pro­duce longer per­for­mance sequences and eval­u­at­ing how coher­ent or context-appropriate they sound in extend­ed impro­vi­sa­tion might reveal new facets of pat­tern syn­er­gy that short exam­ples do not expose.

Cross-Cultural Applicability

The strate­gies deployed here for tabla might prove adapt­able to oth­er deeply mnemon­ic or impro­visato­ry musi­cal tra­di­tions (e.g., West African drum­ming, Middle Eastern per­cus­sion). Investigating how the mod­el gen­er­al­izes across cul­tures could under­score the method’s ver­sa­til­i­ty and reveal new limitations.

Conclusion

By merg­ing for­mal lan­guage the­o­ry with eth­no­mu­si­co­log­i­cal field­work and machine learn­ing, the authors pro­pose a pow­er­ful mod­el for cap­tur­ing core aspects of tabla impro­vi­sa­tion. The frame­work encour­ages close human–computer col­lab­o­ra­tion through dynam­ic ques­tion­ing and incre­men­tal gram­mar build­ing. This approach not only advances a cognitive-computational per­spec­tive on music but also opens a path­way for fur­ther inquiries into cross-cultural appli­ca­tions, time-sensitive per­for­mance mod­el­ing, and cre­ative com­po­si­tion with­in implic­it musi­cal grammars.

Download this paper