A COMPLEX SIGNAL ICA BASED ON UNITARY TRANSFORMATION AND COMPLEX HERMITE MOMENT FOR SPEECH SEPARATION

摘要

Recently, the problem of independent component analysis and/or blind signal separation becomes a very popular and emerging field of research, because the problem contains many potential applications, e.g., speech/image enhancement and/or recognition, noise reduction and so on. However, there are problems in a case when observations have time differences between source signals and are affected by system (room etc.). This paper treats, especially, the mixed speech influenced by the time difference or acoustic system. The blind separations for real signals and for complex signals are summarized, both of which have been previously proposed. The mixed voices are first narrow-banded and standardized, and then blind separation is achieved in a frequency domain by maximizing the evaluation criterion composed of complex Hermite moments. Further, the uncertainty in the permutation of separated signals is overcome by using the correlation coefficient between the temporal standard deviation of separated signals, and that in the magnitude of separated signals is solved from the linear relation between the observed and separated signals.