Blind Separation of Time-Delayed and Convolved Sources
This chapter considers the multichannel blind deconvolution problem. Blind deconvolution refers to the problem of determining the impulse response of a system where the output is usually accessible and the system as well as the input are inaccessible. Multichannel blind deconvolution refers to the fact that multiple channels are observable and multiple sources are mixed and convolved simultaneously.
KeywordsSpeech Signal Finite Impulse Response Finite Impulse Response Filter Blind Deconvolution Automatic Speech Recognition System
Unable to display preview. Download preview PDF.
- 1.Its name is due to statistics of the deconvolved signal which are approximately Bussgang (Bellini, 1994). The Bussgang statistic refers to Bussgang at Bell Labs who found that the autocorrelation and the correlation between the signal and its nonlinearly transformed signal exhibit the same characteristics.Google Scholar
- 2.A model that can measure the power spectrum with a pole-zero transfer functionGoogle Scholar
- 3.The natural gradient is also valid in the multichannel case (Amari et al., 1997b).Google Scholar
- 4.i.e. microphones that can record signal sources witch may be far away from the microphone.Google Scholar
- 5.These audio-files are available in http://www.cnl.salk.edu/r,tewon/.Google Scholar
- 6.Fundamentals in speech recognition are presented by Rabiner and Juang (1993) and Deller et al. (1993).Google Scholar
- 7.The speech recognizer is trained with speech signals obtained from various people reading parts of the Wall Street Journal.Google Scholar