A modified LEGION using a spectrogram for speech segregation

The neurobiological approach for auditory scene analysis (ASA) is discussed. We deal with Wang and Brown's model (1998) (W & B model) and describe an experimental study of signal processing by a neural oscillator network model. Wang and Brown claimed that their model is consistent with various neurobiological findings. However, this model is actually inadequate as a neurobiological model in some points, for example, its assumption of delay lines. We attempt to obtain a new interpretation of their oscillator network model from the viewpoint of signal processing. It is interesting to reconsider this method from a different point of view. Thus, we regard their LEGION segmentation as a threshold filtering or an exploration algorithm of a time-frequency plane by replacing the biological-motivated preprocessing with standard signal processing. The proposed model consists of this segmentation and a simple grouping algorithm based on a standard pitch extraction method. Experimental results show that the proposed model can, in some cases, resolve more elaborate harmonic-sound segregation than can the W & B model. In the final section, problems in the results and the general biological-motivated approach are discussed.