[Smt-talk] RN analyzers

Phillip Kirlin pkirlin at cs.umass.edu
Sat Feb 25 17:31:21 PST 2012

One goal which this sort of analysis can contribute to is a 
computational model of how humans understand and process music.  It is 
plausible that having a Roman numeral analyzer will allow us to improve 
other music-related algorithms, like those for music recommendation or 
algorithmic composition.  Having such a program would also be incredibly 
useful for intelligent tutoring systems for music theory, or for 
integrating into music composition (e.g., scorewriting) software.

Following up on Michael Cuthbert's comments, having a corpus of analyzed 
music will also enable further studies to see if patterns in functional 
harmonic analysis are related to, for example, patterns in expressive 

Also, from a computer science perspective, the technical challenges of 
solving such a problem are not to be overlooked.  Studying problems like 
this may lead us to solutions that can be generalized to solve other 
problems.  We may obtain new ways of modeling computational problems, or 
for performing learning or inference using those models.


Phil Kirlin
PhD Student
University of Massachusetts

On 2/25/12 11:10 AM, Michael Scott Cuthbert wrote:
> Dear Christian,
> The output of programs to automatically analyze pieces are usually far
> less important than the class of questions that having large, searchable
> repertories of analyzed pieces opens up. So for instance one could ask
> “Does V43 move to I or I6 more often during the period 1700-1900 and how
> does the distribution of resolutions change according to time period,
> region, and/or composer?” If the software is more trustworthy then you
> can ask rather finely grained questions; if the software is less
> accurate then you can probably still find things like whether the
> relative frequency of a common chord differs significantly between one
> area and another.
> One thing that the work of De Clercq and Temperley show for me is that
> there’s a limit to the accuracy of any system, somewhere around 95% --
> beyond that even trained professionals will disagree about the labeling
> (even taking into account orthography differences such as I64 vs V64) of
> chords; going above 95% most likely means that the system has learned to
> mimic one analyst’s labels, not that it’s getting the “true” label for
> any given chord.
> Best,
> Michael
> *From:*Christian Goursaud [mailto:christian.goursaud at gmail.com]
> *Sent:* Friday, February 24, 2012 17:24
> *To:* Michael Scott Cuthbert
> *Cc:* smt smt-talk
> *Subject:* Re: [Smt-talk] RN analyzers
> I would be interested to hear what are the proposed goals of the
> development of such software, beyond the obvious technical challenges
> posed to developers. What would be the ultimate utility of software
> which could mimic conventional human translation of music-notational
> data to roman-numeric functional analysis?
> Best,
> Christian
> --
> Christian Goursaud
> PhD candidate, Birmingham Conservatoire
> +44 (0)7796033585
> christian.goursaud at gmail.com <mailto:christian.goursaud at gmail.com>
> _______________________________________________
> Smt-talk mailing list
> Smt-talk at lists.societymusictheory.org
> http://lists.societymusictheory.org/listinfo.cgi/smt-talk-societymusictheory.org

More information about the Smt-talk mailing list