A two-stage deep learning framework for lead instrument recognition in polyphonic music featuring Chinese instruments

Area of Science:

• Artificial Intelligence
• Music Information Retrieval
• Computational Musicology

Background:

• Deep learning has advanced music analysis, but struggles with traditional Chinese music's complexity.
• Discerning lead instruments in polyphonic Chinese music is underexplored, with current models failing on real-world recordings.

Purpose of the Study:

• To develop a deep learning framework for accurate Chinese instrument recognition in polyphonic music.
• To address limitations of existing models in handling complex acoustic features of traditional Chinese music.

Main Methods:

• A two-stage "separation-then-classification" deep learning strategy was employed.
• A source separation module extracts individual instrument audio, followed by a multi-label classification network for identification.

Main Results:

• The framework significantly enhanced accuracy, precision, recall, and F1 scores on a new dataset.
• Notable improvements were achieved in distinguishing instruments with similar timbres, like dizi and xiao.

Conclusions:

• The proposed method offers a scalable, model-agnostic approach for polyphonic instrument recognition in traditional music.
• This work provides foundational technology for preserving and analyzing Chinese musical heritage using intelligent audio systems.