This is the page where you can make program suggestions for the ISMIR 2013 "Demos and Late-breaking" track

(go here for more general information about the ISMIR 2013 Demos and Late-breaking track)

Put your program suggestions below. If a category is missing, feel free to add it. Don't forget to link your name to your email or twitter account!

1) How to edit?
  • Simple: click on the "edit" button on the top right corner of this page, type away, and save. Upon submission, you'll be asked to complete a quick form. Be sure to save your work and check back as simultaneous edits cancel each other!
  • To propose a topic/session, put a tentative title in bold, then a quick paragraph explaining what you're planning. End it with "proposed by [yourname]" with a link to your email, twitter or some other way we can contact you.
  • To comment on a session (add up to its scope, say you're interested, etc.), add a bullet point after the proposal's main paragraph (see examples below). End your comment with your name, as above.
  • Have a look at last year's program suggestions
  • If you have any questions or problems, ask Mohamed Sordo for help.


2) Structure

Topic title

....little description of the talk... (by link_to_author's_email_or_twitter)
  • ...comment (interest, discussion, etc.)...(by link_to_author's_email_or_twitter)
  • ...another comment...



Suggestions for this year:


1. Creative Applications of MIR

[ Eric J. Humphrey, @ejhumphrey ]

Like many in the ISMIR community, the vast majority of my research is heavily biased toward pattern recognition tasks, e.g. chord identification, instrument classification, auto-tagging, and so on. While the ends in and of themselves are well motivated --help humans navigate and search volumes of recorded music-- recently I find myself less interested in listener-facing applications of MIR, and more in the creative uses of musically intelligent systems. And with an eye to the future, all forms of media are becoming increasingly interactive, shifting emphasis from “users-as-consumers” to “users-as-producers.” Noting that music has always been a read-rewrite cultural process, I think our community is positioned to capitalize on a huge opportunity.
While I’m aware that this is hardly a new research area, I think it would be useful --especially now-- to have a comprehensive discussion of this space. Therefore, I would like to convene a session to survey the state of the art in creative applications of MIR and address some of the following:
- A brief review previous work
- Clustering into main topics
- Developing a unified research trajectory
- Evaluation best practices, or “How on earth do we assess creativity?”
- Near-field challenges

I’m content to moderate the discussion, but would generally solicit help / input from anyone interested. Also, more topics from those more informed than myself would be great as well!

Comments:
  • I think this is a fantastic suggestion and hope to discuss this topic in relation to source separation and music remixing applications. Thanks! Nicholas J. Bryan

  • "Creative Music-IR" sounds like a fantastic topic! I'll be giving a talk at the upcoming NIPS workshop on "Constructive Machine Learning" in December (2013). It would be great to learn more about what others are currently doing within the Music-IR community so that I can share our work with a more general machine learning audience. Doug Turnbull

2. Well-Defined Tasks and Good Datasets for Large Scale MIR

[ Eric Battenberg ]

In an attempt to deal with the complex nature of music signals, the models and techniques used by MIR researchers are becoming increasingly complicated. In order to adequately evaluate such techniques and to justify the increasing complexity, it is important to make sure that techniques are evaluated in a thorough manner which is consistent with their intended real-world use, and that the datasets used in training, validation, and testing are well curated, representative of the real world, and of sufficient size.

Other related disciplines such as computer vision and speech processing have addressed these concerns with very specific tasks tested in evaluation competitions on large, publicly available datasets. The MIREX initiative has been a great resource for these comparative evaluation and dataset curation needs for the MIR community; however, the standardization of datasets and tasks has been less apparent in published work. With the exception of the oft-used (and insufficient) GTZAN genre dataset, many MIR researchers must resort to collecting datasets specific to their particular work, which solves the problem of evaluating their particular technique, but makes comparison with other related methods difficult. Music copyrights obviously play a significant role in limitations in dataset curation.

I'd like to have a discussion about these issues and possibly decide on some directions to improve consistency in evaluation of certain popular tasks and availability of useful datasets. Feel free to discuss this here on the wiki, so I can see how others feel.

Comments:
  • Though I'm not in Curitiba, I'd like to upvote this topic as I think it is of utmost importance for MIR research. It's been acknowledged for too long now, and some recent threads in the music-ir list proved that people are willing to push this forward, so ho ahead! Julián Urbano
  • I totally agree with Eric and Julián here. I think this is one of the most important topics the MIR community should be discussing, if we are really willing to push forward our research area (as a community). I'm really looking forward to attend this discussion at the LB session. Mohamed Sordo

3. Automated feature extraction over distributed datasets

[ Alastair Porter ]

As part of the development of our Dunya application (to be presented this year) we have been working on an automated system that extracts mid-level and high-level features from audio collections.
Feature extractors can currently be written using Essentia's python library, or provided as standalone executables, however this could be extended to support other services too (vamp, matlab executables, java). Extractors are versioned and the output of each version of the extractor is stored so that results between versions can be compared. We make extracted information available over a basic REST webservice.
We are interested in discussing the suitability of a service like this for the wider MIR community, especially in light of recent discussions on the music-ir list about datasets and feature extractors. Some points of discussion could include:

  • How to combine a network of worker machines over many research groups so that researchers can submit an algorithm in a single place and have it automatically run on many different audio collections without having direct access to the audio
  • Can research groups restrict access to their audio collections so that algorithms have access to the audio but people don't?
  • Can there be tit-for-tat accounting to ensure that no group unfairly uses the resources of another group without recompense?

4. Digital Musicology and MIR
[Frans Wiering]

Each year many exciting new technologies are presented at ISMIR. Quite a few of these have a potential to generate new insights in music, whether from a psychological, cultural or musicological viewpoint, in other words to contribute to the growing field of Digital Musicology. I would like to propose a session in which we discuss the relevance of this year’s papers to music research, talk about concrete application scenarios, new projects and collaborations of any size that might are around these, and so on. But specifically I would like to discuss what the current technical needs of the music research communities are. How we can translate these into MIR research challenges that we could hope to see solutions for in the next few ISMIRs?

This session is proposed by Frans Wiering (f.wiering@uu.nl). I am the Chair of the International Musicological Society Study Group in Digital Musicology. The aim of the Study Group is to provide an interface between musicology and technical research in sound and music technology, music informatics and music information retrieval. This session is an activity of the Study Group, but it is open to anyone with an interest in Digital Musicology.

5. MIR in Music Education
[Srikanth Cherla]

I would like to propose a discussion on how various MIR techniques have been/can be applied to music education. Sometime ago I came across the Songs2See project from Fraunhoffer IDMT at the Music Tech Fest held in London this year. It was very interesting to see how using MIR, music education can be tailored to individual students' needs and interests, with the possibility of making the learning experience engaging. Another project which focuses on making music education more accessible, while at the same time making it easier for teachers to attend to many students is PRAISE. A related initiative in this regard which deals more generally with education is the recent NIPS workshop on Personalizing Education with Machine Learning.

My aim in proposing this discussion is to learn more about how MIR can be applied to shape music curricula, to make music education more accessible to everyone, particularly those that do not have easy access to music schools, or cannot afford to enroll in full-time courses at universities due to limited availability of time or funds. I would like to hear the opinions of those that have experience in teaching music, or have studied music at university about different ways in which MIR can be incorporated into music education, and the possible benefits of this. I welcome any comments/links related to this subject that will help the discussion.

Tentative pathway for discussion:
  • A review of state-of-the-art in MIR applied to music education.
  • Reflection on shortcomings of previous projects in this area.
  • How have recent MIR developments (for example, deep learning, culture-dependent MIR, computational creativity) opened up new possibilities?
  • Differences in music teaching approaches influenced by culture/region.
  • Relating more general applications of AI, data-mining, signal processing, cognitive science, psychology and software engineering to education (see additional links below).
  • The desired extent of the involvement of MIR in music education.
  • Specific learning tasks (composition, improvisation, etc.) that can benefit from MIR systems.

Some additional links:

6. Music Segmentation: Perception, Evaluation and Inherent Challenges
[Oriol Nieto, @urinieto]

At this year's ISMIR, the task of music segmentation has had a significant presence, leading to some really good discussions about it. I am particularly interested in how we, as humans, perceive the structure of a music piece, and how we can apply this knowledge to (i) design better systems to automatically segment and label music, and (ii) improve the evaluation metrics that judge these systems. After last year's success of the late breaking session on music segmentation, I would like to propose a follow up, and discuss the latest advances in the field.

Some possible discussions (in no particular order):
- Boundaries evaluation: why 3/0.5 seconds window? Would it make more sense to have a "weighted" window instead?
- Labels evaluation: Pair-wise clustering vs entropy values (vs random? vs ...?). Which one is more reliable? What are the pros/cons of each?
- Deep Learning: Can we learn better structural features automatically? Most of the publications on music segmentation either use MFCC or Chromas (and when they combine them, it's usually simply by stacking one on top of the other). Could we use deep learning to learn a representation that captures both timbre and harmony (and maybe rhythm and... melody lines?) in an optimal way for music segmentation?
- Perception and functional structure: is there more room to use perception and cognition findings in segmentation tasks?
- Vice versa: can our annotated data be used to address questions in the perception of boundaries and structural function. What data would be needed to answer more of them? (@jvanbalen)

7. Here's an additional list of topics or ideas that I'd like to see discussed at the D&LB session this year.
  • Linked Data in MIR, ontologies.
  • Increased usage of MusicBrainz (open metadata) in recent publications @ ISMIR
  • Definition and perception of Melody in different musical cultures.
  • What can we really learn from non-western art music traditions and apply it in Western music?
  • Personalization, Visualization & User Experience.
  • Journal of the International Society for Music Information Retrieval (JISMIR). Selected papers of the conference (based on votes by the program committee, probably) get the chance to be extended and published at this Journal.
If you have any more ideas, please feel free to add them here. If you are really interested in any of these topics, please also feel free to create a new section for it and extend it. Mohamed Sordo
  • About the JISMIR, it's regular to select papers of the conference. Is it possible to combine the results of MIREX contest and submissions of ISMIR (not just conference paper), which could have some late-breaking of their research, but not ready on the time frame of ISMIR submission and review (it almost has 6 months different from submission to conference holding.)? The factors of consideration could be not published with certain period from beginning of the research topic and have good result on MIREX with first priority. Frank Wu

8. Realizing Semantic Web Technologies in Music Digital Libraries

A digital library is defined as: afocused collection of digital objects—including text, audio and videoalong with methods for access and retrieval, and for selection, organisation, and maintenance. Recent work at Illinois, McGill Oxford, and Waikato Universities has shown how semantic web technologies can be integrated into this definition, using music digital libraries as a motivating example. The end result goes beyond traditional expectations as to what this type of digital resource can provide. Two specific realizations of this work are: i) the SALAMI Linked Data Resource (including a Sparql endpoint) providing a gateway to over 2.1 million music segmentation files; and ii) Unchained Melody -- a sequence of musical digital libraries that demonstrates how semantic web technologies can change the services provided by a digital music library.

This work draws heavily upon linked data, makes use of MusicBrainz and so aligns strongly with item (7) above. David Bainbridge., J. Stephen Downie



Demos:

Demonstration of Dunya - a web app for exploring music exploiting cultural context

[ Alastair Porter ]

During ISMIR we are presenting a poster on our current development of the Dunya application, designed to explore the music of the south of India based on research performed as part of the CompMusic project. Dunya utilises some novel methods for making recommendations based on specific knowledge of the musical style and the culture that it is a part of. If there is additional interest we are happy to discuss in more detail some of our design decisions, architectural information, and features used in the application.


Demonstration of ISSE - An Interactive Source Separation Editor


We've been working on a new audio editing tool called (ISSE) to perform source separation using drawing and painting tools (alpha-release). We would like to show a few examples, and then discuss an in-the-works third-party plugin API to allow others to leverage the same interface of new and alternate separation algorithms. It would be great to hear what other researchers would like for such a development platform and get general feedback on the software.

Nicholas J. Bryan, Gautham J. Mysore, and Ge Wang.

Demonstration of Essentia - An open source C++/Python library for audio analysis and music information retrieval


http://essentia.upf.edu

We present Essentia in a poster session on Thursday. In this demo session we would like to show more details about Essentia and how you can use it for your applications and research purposes.

Essentia is an open-source C++ library with Python bindings for audio analysis and audio-based music information retrieval released under the Affero GPLv3 license (also available under proprietary license upon request). It contains an extensive collection of reusable algorithms which implement audio input/output functionality, standard digital signal processing blocks, statistical characterization of data, and a large set of spectral, temporal, tonal and high-level music descriptors. In addition, Essentia can be complemented with Gaia, a C++ library with python bindings which implement similarity measures and classifications on the results of audio analysis, and generate classification models that Essentia can use to compute high-level description of music (same license terms apply).


Demonstration of Spot the Odd Song Out and Casimir - API and Framework for data collection with on-line games


The Casimir framework supports the development of data collection by on-line games (or surveys), which run inside a web browser. We show the "Spot the Odd Song Out" game which consists a set of Casimir game modules, which collect data on audio music similarity, meter, and rhythm. The Casimir API supports storage, organisation of data and selection of tasks to achieve a desired properties of the dataset (distribution, connectedness etc.). The modules and the framework are openly available to support more data gathering and sharing in MIR.

Daniel WolffGuillaume BellecTillman Weyde


Interactive Demonstration of Performance System from "Elvis by Elvis" Concert Piece


Elvis by Elvis was one of the selected pieces for Concert II at ISMIR 2013. The singer performed with a modular system which connected multiple MIR-based tools and programming languages to create a realtime, non-deterministic musical performance. The singer sang Elvis Costello melodies, and was accompanied by a sequence of short audio segments of Elvis Presley songs, aggregated based on the underlying chord harmony present over that segment. The system combined 3 main modules- a realtime pitch-tracker built in Supercollider, a realtime chord progression generator fuelled by a Hidden semi-Markov Model trained on David Temperley's Rolling Stone Rock 'n' Roll Corpus and built in C++, and a Chuck-based audio management system which played audio labelled by the generated chord (using labels from the Billboard dataset). Each system communicated using OSC messages over UDP. The system will be live, so anyone can sing in a melody and automatically generate an Elvis Presley accompaniment in realtime.

Ryan Groves Thor Kell