New Frontiers in MIR

A talk about the opportunities that new machine learning models and datasets present to the field of music information retrieval.

Transcript

1. “I DON’T KNOW ANYTHING ABOUT MUSIC. IN MY LINE, YOU DON’T HAVE TO.” -ELVIS NEW FRONTIERS IN MIR
3. MUSIC INFORMATION RETRIEVAL ▸ The way we approach studying music in a digital context is totally different from how we normally think about music ▸ When processing an audio file, it’s amazing enough that we can take a waveform and recreate complex sounds from it. How do we extract more meaningful information from this?
4. THE STATE OF MIR ▸ There are a few major problems that have guided development in the field for as long as it’s been a thing ▸ Example: Pitch recognition and the problem of overtones ▸ Example: Transcription and the difference between how a song might be performed and how it should be transcribed
5. THE FUTURE OF MIR ▸ We are seeing the early stages of an explosion in MIR and computational musicology, and combined with the increasing accessibility of machine learning and new music datasets (not to mention advances in other relevant fields), it will only get cooler ▸ MusicNet
6. STUDYING MUSIC ACTUALLY ISN’T A WASTE OF TIME (REALLY) ▸ This huge-scale study of music is unprecedented, and will have major impacts on other fields ▸ Existing medical studies can be combined with insights from MIR & computational musicology — e.g. the already existing and pretty amazing studies about processing pain, the impact of music education on speech and language development, the relationship between music education and neuroplasticity, etc. ▸ Very few processes in everyday life use as many discrete parts of the brain as playing/processing music — understanding it inevitably means understanding more about the brain, too
7. ▸ We could actually almost think about: ▸ Non-invasively scanning a person’s brain to identify the subjective, tiny musical attributes that most ease pain || fight anxiety/depression || motivate || whatever, really, + ▸ Training a machine learning algorithm based on that user’s brain imaging data, + ▸ Using that program to generate new music that can not only (e.g.) reduce the amount of painkillers necessary for surgery (above study suggests their relatively un-curated playlist can already do this by about 20%), but keep getting better at it too