CUMIN are excited to announce that our first project is underway!
Music Transcription with Deep Neural Networks
Starting off with solo piano pieces, develop a machine learning solution that automatically transcribes a music recording into a score.
There has been some work on this done by Google’s Magenta team https://magenta.tensorflow.org/onsets-frames . Their solution works fairly well, but is far from perfect. However, they have used a pre-cured dataset of transcriptions and piano recordings.
The project would aim to improve their solution by using software generated data; by utilising any of the wide array of virtual instruments and synthesizers one could generate large amounts of realistic, varied and labelled audio data automatically. Furthermore, many different recording scenarios can be simulated through data augmentation with reverb, mic and room-simulator and other effects.
The project leader has emailed the author of the Google Magenta’s paper, and he said they haven’t considered this approach extensively and thinks it could work.
If successful, this could be extended into transcription of other instruments and transcription from recordings with noise (e.g. transcription of piano riff from a song with other instruments).
Team leader’s email: email@example.com