Micropower Mixed-signal VLSI Independent Component Analysis for Gradient Flow Acoustic Source Separation

Area of Science:

• Mixed-signal VLSI design
• Acoustic signal processing
• Machine learning algorithms

Background:

• Independent Component Analysis (ICA) is crucial for separating mixed audio signals.
• Existing ICA implementations often lack the miniaturization and power efficiency required for portable devices.
• Reverberant environments pose significant challenges for accurate source localization.

Purpose of the Study:

• To develop a parallel, micro-power mixed-signal VLSI implementation of ICA.
• To enable real-time separation and localization of multiple sound sources.
• To create a compact and energy-efficient microsystem for acoustic sensing applications.

Main Methods:

• Utilized gradient sensing of the acoustic field with a miniature microphone array for pre-processing.
• Implemented reconfigurable outer-product learning rules for ICA.
• Designed and fabricated the processor using 0.5 µm CMOS technology.

Main Results:

• Successfully separated and localized up to 3 sound sources in mildly reverberant conditions.
• Achieved comparable performance to benchmark FastICA algorithms, with enhanced robustness in noisy environments.
• Demonstrated perceptually clear separation and precise localization of speech sources.

Conclusions:

• The developed ASIC offers an extremely small form factor and low power consumption for source separation and localization.
• The microsystem is suitable for integration into intelligent hearing aids and wireless distributed acoustic sensor arrays.
• This work advances the feasibility of advanced acoustic processing in resource-constrained devices.