The objective of this research is design of algorithms and real-time system implementations for task of source localization and separation using miniature sensor arrays, where dimensions of the arrays are much smaller than wavelength of the incident signals. The approach is the integration of wavefront sensing and independent component analysis in unique framework that extends the performance of current source separation algorithms and lends itself into efficient implementation in mixed-signal VLSI, yielding real-time performance at small-form factor.

Gradient flow is a signal conditioning technique for arrays of very small dimensions, which converts time delays between signal observations into relative amplitudes of the time-differentiated signal, by observing gradients (spatial differences). Improved differential sensitivity of gradient sensing allows to shrink the aperture of the sensor array without degrading signal-to-noise ratio. Besides its use in bearing estimation, gradient flow greatly simplifies the the problem of separating unknown delayed mixtures of independent signal sources by transforming it to a simpler problem of separating corresponding instantaneous mixtures of the time-differentiated signals.

A mixed-signal VLSI system comprising sensing of the spatial and temporal differences (gradients) of the acoustic field at very small aperture and static independent component analysis was designed. The real-time performance of the system demonstrates perceptually clear (12dB) separation and precise localization of two speech sources through experiments in a conference room setting. Miniature size of the microphone array enclosure (1 cm diameter) and micropower consumption of the VLSI hardware (250 uW) are key advantages of the approach, with applications to hearing aids, conferencing, multimedia, and surveillance. As example of the separation performance of the gradient flow, the following speech signals were recorded in a conference room using 4 hearing aid microphones: x10, x01, xm10 and x0m1.
The recovered source signals by gradient flow ICA: Recovered source 1 and Recovered source 2

For source separation in reverberant environment, a subband decomposition of the spatial and temporal gradients, with static ICA applied in each frequency band, is proposed. The localization results obtained from the ICA applied on the unfiltered gradients resolve the scaling and permutation indeterminacy inherent to the static ICA and improve the separation performance in moderate reverberant environment. The devised subband ICA algorithm with sensor array consisting of 8 microphones improves the separation performance by 3 dB in the case of the moderate reverberant environment.

We have proposed the architecture of the microsystem for the source separation in the room environment under moderate reverberation. The architecture comprises 16-channel filter bank with the static linear independent component analysis implemented in each subband. In the proposed implementation, preserving the linear mixing of the source signals in the ICA model is critical for the source separation. This is the reason why in the proposed microsystem the requirement on the linearity of the filters is stringent. Although, the gm-C filters don't achieve linearity that can be achieved by switched-capacitor filter implementation, the choice of the gm-C filter implementation is dictated by the chip area and the power consumption requirements. We have proposed a new implementation of the auditory filter bank that meets the linearity requirements with the large dynamic range. The proposed microsystem with the subband decomposition and static ICA in each subband has demonstrated 13 dB separation under moderate reverberations (reverberation time of 300ms) with signal-to-interference (SIR) improvement of the 3 dB over the architecture with the separation of full-band signals.

S. Li, Y. Lin and M. Stanaćević, "Mixed-signal VLSI Microsystem for Acoustic Source Separation", Proc. 56th. IEEE Midwest Symp. on Circuits and Systems (MWSCAS'2013), Columbus, Ohio, 2013.

Y. Lin and M. Stanaćević, "A Low-Power, High-Linearity Filter Bank for Auditory Signal Processing Microsystem", Proc. 56th. IEEE Midwest Symp. on Circuits and Systems (MWSCAS'2013), Columbus, Ohio, 2013.

S. Li and M. Stanaćević, "Subband Gradient Flow Acoustic Source Separation for Moderate Reverberation Environment", Conf. Rec. of the 46th Asilomar Conference on Signals, Systems and Computers, Pacific Grove CA, Nov 2012.

S. Li and M. Stanaćević, "Gradient Flow Source Localization in Noisy and Reverberant Environments", Conf. Rec. of the 46th Asilomar Conference on Signals, Systems and Computers, Pacific Grove CA, Nov 2012.

A. Chacon-Rodriguez, S. Li, M. Stanaćević, L. Rivas, E. Baradin and P. Julian, "Low Power Switched Capacitor Implementation of Discrete Haar Wavelet Transform", Proc. 3rd. IEEE Latin American Symp. on Circuits and Systems (LASCAS'2012), Feb 2012.

S. Li, X. Yun and M. Stanaćević, "Low-power System-on-chip Acoustic Localizer", Proc. 53rd. IEEE Midwest Symp. on Circuits and Systems (MWSCAS'2010), Seattle, WA, August 1-4, 2010.

Recordings from the four microphones with the array diameter of 1cm are available to the research community. Please contact us for the access to the full database.