Robust Real-Time Music Transcription with a Compositional Hierarchical Model

Area of Science:

• Music Information Retrieval
• Machine Learning
• Signal Processing

Background:

• Traditional music transcription methods often struggle with accuracy and real-time processing.
• Hierarchical structures are inherent in music, suggesting their utility in transcription models.

Purpose of the Study:

• To present a new compositional hierarchical model for robust music transcription.
• To enable insights into the learned representation through transparency.
• To achieve robustness and speed for real-world, real-time applications.

Main Methods:

• Developed a multi-layered compositional hierarchical model.
• Employed unsupervised learning to build hierarchical representations from music signals.
• Utilized statistical co-occurrences to compose parts across layers.

Main Results:

• The model demonstrates robustness and speed suitable for real-time use.
• Evaluated performance in multiple fundamental frequency estimation.
• Hierarchical structure aligns well with musical hierarchies, from partials to chords.

Conclusions:

• The proposed model offers a transparent and effective approach to music transcription.
• Its unsupervised, hierarchical nature makes it adaptable for various music information retrieval tasks.
• The model shows promise for advancing automated music analysis and transcription.