|
Applied Math Seminar
Time and Frequency Domain Methods of Blind Source Separation
|
|
Blind source separation aims to recover source signals from the received convolutive mixtures in an unknown environment. Nonlinear algebraic equations and associated minimization problems are derived from statistical independence of source signals. However, solutions are not unique, and minimization is non-convex. Additional selection criteria are introduced based on properties of room impulse responses, times scales of signal spectral dynamics, and L1 minimization. Two recent time and frequency domain methods will be discussed and results on recorded speech/music signals will be demonstrated. |