Unsupervised Learning for Monaural Source Separation Using Maximization⁻Minimization Algorithm with Time⁻Frequency...

Area of Science:

• Signal Processing
• Machine Learning
• Audio Analysis

Background:

• Nonnegative matrix factorization (NMF) is crucial for signal decomposition.
• Traditional NMF methods often use fixed divergence measures (e.g., Kullback-Leibler, Least Squares).
• Fractional β-divergence offers a more flexible cost function for NMF.

Purpose of the Study:

• To develop an unsupervised learning algorithm for sparse nonnegative matrix factor time-frequency deconvolution.
• To optimize the deconvolution process using a generalized fractional β-divergence.
• To improve source separation performance in audio mixtures.

Main Methods:

• An unsupervised learning algorithm based on fractional β-divergence was developed.
• A maximization-minimization (MM) algorithm was employed for fast convergence.
• The algorithm decomposes time-frequency representations into spectral dictionaries and sparse temporal codes.
• A method for estimating the fractional β value was proposed.

Main Results:

• The proposed algorithm demonstrated efficient extraction of spectral dictionaries and temporal codes.
• Optimized deconvolution led to significantly improved audio source separation.
• The method showed superior performance compared to existing factorization techniques.
• Guaranteed convergence was achieved through the multiplicative update algorithm.

Conclusions:

• The generalized fractional β-divergence provides a powerful tool for sparse nonnegative matrix factorization.
• The developed unsupervised algorithm offers enhanced efficiency and accuracy in time-frequency deconvolution.
• This approach significantly improves performance in single-channel audio source separation tasks.