New self-normalized blind source separation networks for instantaneous and convolutive mixtures
✍ Scribed by Yannick Deville; Ovidiu Albu; Nabil Charkani
- Publisher
- John Wiley and Sons
- Year
- 2002
- Tongue
- English
- Weight
- 120 KB
- Volume
- 16
- Category
- Article
- ISSN
- 0890-6327
- DOI
- 10.1002/acs.722
No coin nor oath required. For personal study only.
✦ Synopsis
Abstract
The main blind source separation networks proposed in this paper apply to convolutive mixtures (including instantaneous ones). They have a recurrent or direct structure and they may use channel‐specific separating functions. They are based on a self‐normalized weight adaptation rule, which adaptively estimates the average powers of non‐linear functions of the network outputs. This allows us to control several aspects of the operation of these networks, especially their convergence speed/accuracy trade‐off. It also makes them more robust with respect to non‐stationary situations. We analyse their convergence properties. We validate all these results by means of experimental tests performed with these networks, classical ones, and additionally proposed linear instantaneous direct networks based on a normalization of their outputs. These tests especially show that the proposed networks improve the convergence trade‐off and that only these networks apply to highly mixed non‐stationary sources. Copyright © 2002 John Wiley & Sons, Ltd.