Authors:

Elan L

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Music Transcription is a process of converting audio into a symbolic representation. Automatic Music Transcription (AMT) promises to enable improved accessibility for musicians as well as enabling and superseding other technologies. Several methods have been developed in the field of AMT. This research explores these different approaches. The approaches that we found that show the most promise for AMT are identified along with their challenges and limitations. Machine learning (ML) shows promise among AMT schemes due to the room for growth while achieving state-of-the-art AMT performance. ML techniques such as sequence-to-sequence transformers and convolutional neural networks have been found particularly effective for performing AMT. Issues with these techniques include context length limitations, imprecision, inaccuracy, and large data requirements. We identify some techniques applied to ML in other domains, such as natural language processing. We repurposed these techniques to overcome some of the challenges encountered when applying ML to AMT. New ML architectures are identified that enable improved scaling of ML models to larger contexts. Methods for data augmentation are identified to enable more effective training on small feature sets. Future work will involve applying the identified strategies and comparing AMT effectiveness with the new strategies vs without. Keywords: Automatic Music Transcription; Machine Learning; Music

Source:

Purdue University / 2024

Topics:

No topics listed

Co-authors:

Elan L