# Melody Extraction from Polyphonic Music by Deep Learning Approaches: A Review

**Authors:** Gurunath Reddy M, K. Sreenivasa Rao, Partha Pratim Das  
**arXiv:** [2202.01078](https://arxiv.org/abs/2202.01078)  
**Date:** 2022-02-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2202.01078v1/

## Key points

- Melody extraction is crucial in music information retrieval for applications in education and the music industry.
- Extracting melodies from polyphonic music is challenging due to background instruments with similar characteristics to the melodic source.
- Recent advancements in deep learning have led to data-driven approaches being explored for melody extraction.
- The paper reviews current deep learning techniques for melody extraction, categorizing models based on neural network type and output representation.
- It provides overviews of 25 different melody extraction models, discussing architectures, loss functions, input representations, and parameter settings.
- The paper includes a section on explaining black-box deep neural networks used in melody extraction.
- Performance comparisons of the 25 models are made, highlighting strengths and limitations.
- Possible future directions include exploring different input representations to enhance model performance.

## Summary

In the field of music information retrieval, melody extraction plays a crucial role for its applications in education and the music industry. The task of extracting melodies from polyphonic music is particularly challenging due to the presence of background instruments that often exhibit similar characteristics to the melodic source. This interference makes isolating the melody from the mixture signal a complex endeavor. While classical signal processing methods were previously popular for melody extraction, recent advancements in deep learning have inspired researchers to explore data-driven approaches. This paper provides a comprehensive review of current deep learning techniques for melody extraction from polyphonic music. The authors categorize available deep models based on the type of neural network and output representation used for predicting melodies. They present brief overviews of 25 different melody extraction models, discussing their architectures, loss functions for model optimization, input representations, and parameter settings. Additionally, the paper includes a section on explaining black-box deep neural networks used in melody extraction. The performance of these 25 melody extraction methods is compared, highlighting their strengths and limitations. Furthermore, the paper discusses possible future directions for improving melody extraction methods, such as exploring different input representations to enhance model performance. The authors suggest experimenting with variations in input representation while keeping the model architecture constant to assess the impact on performance. Overall, this review offers valuable insights into the current landscape of deep learning approaches for extracting melodies from complex musical compositions.