User Tools
**This is an old revision of the document!** ----
====== DeepPerformanceDiff ====== | Authors | [[https://www.twitter.com/utstikkar|Amélie Anglade]], [[https://www.twitter.com/jganseman|Joachim Ganseman]], [[https://www.twitter.com/peachnote|Vladimir Viro]] | | Affiliation | Various | | Code | [[https://github.com/peachnote/sync-visualization|Github Link]] | **DeepPerformanceDiff** allows you to compare renditions of the same piece (of which the score is provided by Peachnote through IMSLP) by different performers (provided by YouTube). Differences in tempo and loudness are visualised in a graph that makes it easy to spot passages that are played markedly different from others. ===== Origins ===== This project originates from an idea that is part of the [[http://music-connection-machine.peachnote.com/|Music Connection Machine]] brainstorm taking place at [[http://www.peachnote.com|Peachnote]]. A minimal mockup version had already been coded. During the HAMR@ISMIR hackday, this mockup has been extended to a fully functional prototype include more audio features (notably loudness), better search functionality, and a more flexible layout. ===== Used Technologies and Data ===== * Most of the code has been written in JavaScript. * Calculation of loudness was done in Python using [[http://essentia.upf.edu/|Essentia]]. * The web interface was wrapped in [[http://getbootstrap.com/|Bootstrap]] * [[http://imslp.org/|IMSLP]] scores are displayed using the [[http://www.peachnote.com/api.html|Peachnote API]] * Videos are gathered from [[http://www.youtube.com|YouTube]] ===== Relation to MIR Research and Potential Impact ===== Study of performance practice, and especially differences between performers, is a research topic that has already garnered some interest in the past (([[http://www.sciencedirect.com/science/article/pii/S0004370205000196|Stamatatos & Widmer 2005]])) (([[http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3842509/|Gingras et al. 2013]])) (([[http://www.mirlab.org/conference_papers/International_Conference/ISMIR%202008/papers/ISMIR2008_240.pdf|Sapp 2008]])). Note that many of these papers focus on (weighted combinations of) tempo and loudness differentiation to distinguish different performers, which are the exact features that DeepPerformanceDiff is able to visualize in a single graph. With this project, we are able to make an intuitive visualization of the data. This does not only create possibilities for data exploration, but is also of practical interest. We think more specifically in the direction of performance practice education, where more and more people upload recordings of themselves playing on YouTube. They are now able to compare themselves to others, to professional recordings, or to earlier versions of themselves, which enables them to track their own progress. ===== Try it out! ===== The code is available for download on [[https://github.com/peachnote/sync-visualization|Github]]. Fork/clone the repository and point a webserver to the project directory. Additionally, you need a few precomputed files - [[https://www.twitter.com/peachnote|contact Peachnote]] to obtain them: * Alignment definitions (json) between IMSLP scores and YouTube videos * Alignment Quality definitions (json) between IMSLP scores and YouTube videos ===== Future Work ===== For the study of performance We could most definitely think towards clustering of the tempo/loudness profiles. On such a large body of data, this could give us information on stylistic relationships between performers, maybe allowing us to detect schools of performance or regional differences in interpretation. Music education practice would most definitely be served by a tool that visualizes the progress of students. Professional performances already appear often on YouTube, and given that we are in the age of Instagram and Vine, we foresee no lack of uploads ofamateur performances in the near future either. The comparisons between performances by people on different levels with different abilities may provide valuable pedagogical insights for tutors.
