Dubey's Research: Source Separation from Neurally Inspired Processing

Mohit Dubey, a New Mexico Consortium (NMC) Research Scientist working in the field of neuromorphic computing, has been making strides in his research on Source Separation from Neurally Inspired Processing (SSNIP). This research is a collaboration with Garrett Kenyon and Austin Thresher, also of the NMC.

SSNIP is a neutrally-inspired method of separating bass, drums, vocals and other instruments from sparse encodings of phase-rich Constant-Q representations of stereo musical data.
 
In this research, sparse encodings are generated from learned features that are tonotopic and divided into spectrally and temporally convolutional dictionaries.
 
This method is inspired by the hemispheric lateralization of human auditory cortex, the section of the brain that processes information received through hearing.
 
As opposed to previous approaches, Dubey utilizes different methods to achieve his results. First, he uses a logarithmic frequency representation. He also preserves the phase information of the original sound. This is because the phase problem is the problem of loss of information concerning the phase that can occur when making a physical measurement. Third, he uses multiple encoding dictionaries to capture both fine spectral and temporal information. Last, he removes the noise from the original sources in order to obtain the best results possible.
 
At this point in time, this exciting neutrally-inspired research is only being applied to musical data. Dubey and his team aim to expand their approach in the future to separate sources in speech, biomedical (EEG, fMRI, etc.) and other types of data.

 

 

© 2018 New Mexico Consortium