• Innovation through disruptive and scalable technology .
  • Cutting-edge AI research .
  • Accelerating innovations in research and service .
  • We strive for (and achieve) excellence! .
  • “SotA” (State-of-the-Art) .
  • Visual demo of research and service innovation
  • Human. Machine. Experience Together .

Participation in 2019 International Society for Music Information Retrieval Conference (ISMIR)

2019.12.05

SK telecom’s AI Center participated in ISMIR, held in Delft, Netherlands from November 3rd to November 10th. Dr. Chang-Hyun Kim of T-Brain attended Tutorial on recent research and Main Conference proceeded for oral and poster presentations of selected papers as well as networking session with conference attendees. In addition, AI Center hosted a Korean scientist night for the participating Korean researchers.

The ISMIR Conference introduced new methods of music AI in the field of music source separation, music transcription, voice synthesis/translation, music generation and recommendation, demonstrated real-time source separation through MIREX (Music Information Retrieval Evaluation eXchange)/Industry meetup and hosted networking events for participants.

At the MIREX Challenge held on the last day of the conference, research on OnsetNFrames based on the music transcription by Dr. Chang-Hyun Kim and KAIST MAC Lab's Yong Sang-un won first place in Note-tracking task mixed dataset field. In this study, we applied audio and symbolic data augmentation based on training data augmentation, neural architecture search and hyperparameter optimization by using Meta Learner from T-Brain, and model ensemble to increase the overall performance.

In particular, for MIDI data augmentation, we combined symbolic data augmentation which transforms the music itself, and mix-up data augmentation which puts weight for audio data and then synthesizes them, rather than using acoustically transformed audio file observed in earlier research. In symbolic data augmentation, we transformed MIDI data through key translation, key change, and tempo change by randomly setting the stochastic variables. In the mix-up data augmentation, we chose the mix ratio of two music sources for synthesis through hyperparameter optimization.

<Mix-up image example of Mel-Spectrum used for actual transcription>


Lastly, we held a Korean scientist night and had networking time with researchers in related fields. Through this event, we were able to share various information about ongoing research from Korea's top schools (Seoul National University, KAIST, Gwangju Institute of Science and Technology, Kangwon National University) and industries, and had an opportunity to draw a roadmap for future research in Music AI of T-Brain.