On Training Targets for Supervised Speech Separation.

Area of Science:

• Signal Processing
• Machine Learning
• Speech Technology

Background:

• Speech separation is often framed as a supervised learning task, typically using deep neural networks.
• The ideal binary mask (IBM) is a common target, known for enhancing speech intelligibility.
• Supervised learning offers flexibility beyond binary targets for speech separation.

Purpose of the Study:

• To evaluate and compare the effectiveness of various training targets for supervised speech separation.
• To determine which targets yield the best objective intelligibility and quality metrics.
• To compare supervised methods against traditional techniques like non-negative matrix factorization.

Main Methods:

• Training supervised learning models with different targets: ideal binary mask (IBM), target binary mask, ideal ratio mask (IRM), short-time Fourier transform spectral magnitude and mask (FFT-MASK), and Gammatone frequency power spectrum.
• Evaluating separation performance using objective intelligibility and quality metrics.
• Comparing results with non-negative matrix factorization and other speech enhancement methods.

Main Results:

• Ideal ratio mask (IRM) and FFT-MASK targets outperform other tested targets in objective intelligibility and quality.
• Masking-based targets generally yield superior results compared to spectral envelope-based targets.
• Supervised speech separation demonstrates clear performance advantages over non-negative matrix factorization and traditional speech enhancement techniques.

Conclusions:

• Ratio mask targets are highly effective for supervised speech separation, offering significant improvements in intelligibility and quality.
• The supervised learning framework provides a powerful and flexible approach to speech separation.
• This research highlights the superiority of modern supervised methods for enhancing speech separation performance.