Given a natural language query and a piece of music in digital staff notation representation, the C@merata task [6] evaluates an algorithm's ability to identify where one or more events specified by the query occur in the music. It is the latest example of a longstanding interest in querying music represented as (or derived from) staff notation. The C@merata task challenges researchers to extend current knowledge in two respects:
1. Accepting a music-analytic query in the form of a naturallanguage string, such as "perfect fifth followed by a D4";
2. Reliably retrieving instances of higher-level musictheoretic concepts from staff notation, such as functional harmonies (e.g., "Ib") or cadences (e.g., "interrupted cadence").
One application of an algorithm that performs well on the C@merata task would be within music notation software, so that students could query and hear/see results for the pieces with which they are working, in order to develop their understanding of various music-theoretic terms.
The Stravinsqi-Jun2014 algorithm that was entered in the C@merata task is embedded in a Common Lisp package called MCStylistic-Jun2014 (hereafter, MCStylistic), which has been under development since 2008 [1]. 1 MCStylistic includes implementations of algorithms from the fields of music information retrieval (MIR) and music psychology [2][3][4][5].
From a natural language perspective, there are two types of queries: compound queries such as "a Bb followed a bar later by a C followed by a tonic triad", and ordinary queries such as "perfect cadence". Stravinsqi checks the query string for compound queries and splits it into N query elements if necessary, e.g., "a Bb" and "a bar later by a C" and "tonic triad".
The piece is converted from its MusicXML format to kern format using the xml2hum script. 2 The kern file is parsed by import functions in MCStylistic to give the following representations, which are referred to as point sets: (1) instrument/staff and clef names at the beginning of each staff; (2) bar numbers where time signatures are specified, together with the number of beats per bar, the type of beat, and the corresponding ontime (incrementing time in staff notation); (3) a point-set representation of the piece, where each point represents a note. The five-dimensional point consists of the ontime of the note, its MIDI note number, its morphetic pitch number [3], its duration in crotchet beats, and its numeric staff number; (4) a point-set representation of the piece with three extra dimensions, one each for articulation, dynamics, and lyrics information; (5) a point-set representation of the piece, where each point represents a notated rest.
Each query element is passed to several sub-functions (e.g., harmonic-interval-of-a, duration&pitch-timeintervals, rest-duration-time-intervals, etc.), along with the appropriate point set(s). For example, the function rest-duration-time-intervals takes a query element, the point set of notated rests, and the point set of instrument/staff and clef names as its arguments, because these three information sources are sufficient for locating rests of specific duration. If a query string is ordinary (contains one element only), then the time windows in the first nonempty sub-function's output are passed to a final function that converts the time windows into the XML format required by the task. For compound queries, plausible sequences of time windows for the component query elements are merged before passing to the final syntax-conversion function.
Figure 1 contains example output for the sub-functions harmonic-interval-of-a, (arrows indicate retrieved harmonic seconds), HarmAn->Roman (functional harmonic labels below each staff), and cadence-time-intervals (three perfect cadences surrounded by black boxes). All three functions involve implementing and extending MIR/musicpsychology algorithms to achieve promising results, especially for the higher-level music-theoretical concepts such as functional harmonies and cadences.
Figure 2 shows recall and precision results for the Stravinsqi algorithm on the 2014 C@merata task. The measure metrics reward an algorithm's output if it is in the same bar/measure as a ground-truth item, whereas the beat metrics require an algorithm's output to be in the same bar and on the same beat as a groundtruth item. The mean category in Figure 4 shows the overall results, with Stravinsqi having recall .91 and precision .46 at the measure level, and recall .87 and precision .44 at the beat level. 3Stravinsqi's strong performance on the first eight of twelve categories (pitch, duration,…, melodic interval) is encouraging, as is the small decrease in recall (.91 to .87) and precision (.46 to .44) with the change from measure-to beat-level granularity. The drop in precision for compound queries is due to over-lenient criteria used to select and combine time intervals for the different elements that comprise a compound query. This can be fixed in future work. For triad labelling, Stravinsqi suffered from an under-labelling issue in two instances, missing two first-inversion triads because the same triad in root position preceded them, and the two triads got one root-position label. The triad and texture categories are somewhat underrepresented in the training and test data, and so more attention ought to be given to these categories in future. Less-than-perfect performance on melodic and harmonic interval questions can be attributed to inconsistencies between the task description/training collection, and test collection.
Algorithms that perform strongly on the C@merata task open up new, interesting potential applications in music education and musicological research. The Stravinsqi algorithm described above is one such strong performer, and has effectively solved seven of the twelve C@merata task categories shown in Figure 4 (pitch, duration, pitch and duration, articulation, voice specific, lyrics, and melodic interval). As for the remaining five categories, future work will involve bug fixes, resolving task inconsistencies, and acquiring more data for cadence and texture query categories. It may also be helpful to have two experts provide annotations for the higher-level music-theoretic concepts. The addition of new, higher-level music-theoretic query categories would be welcome in future iterations of C@merata as well, in order to keep the task at the forefront of research in music computing.