N-path Spectral Decomposition in Acoustic Signals for Sound Identification

The Murmann lab has developed a method for an extraction information from acoustic signals that utilizes low power consumption. N-path filters are used to decompose the original acoustic signals' waveform before downconverting to lower their Nyquist-rate bandwidth. This allows the system to be implemented in the analog domain, and thus only requires low power consumption. Once decomposed, signals can be identified using the SS-SVM model. The SS-SVM model requires 10x less training data than current ConvNet models to achieve a 10% identification error rate, and its fewer parameters prevent SS- overfitting the data. As a result, this design is applicable for low-level sound analysis, including background noise filtering, and could allow other high-powered AI systems to "sleep" until triggered.

Stage of Research