Speech Separation Based On Computational Auditory Scene Analysis

The human auditory system can perceive and locate the sound he focuses mixed with complex environmental sounds. Computational auditory scene analysis tries to model auditory processing by computer technology. We present in this paper a model of speech separation based on this principle to separate a mixture of two sounds and distill target speech. Different from other models, we don't use the same method to separate resolved and unresolved harmonics, so we avoid the problem of inability to separate unresolved harmonics successfully. A new method of pitch evaluation is also presented.Our model is evaluated with a corpus of mixtures, which has been used to test many CASA systems. We also compare our model with Wang-Brown model. The results of experimentation show ours have better performance.