Analog-to-Feature (A2F) Conversion for Audio-Event Classification

Abstract

Always-on sensors continuously monitor the environment for certain events. Such sensors are often integrated on battery-powered devices, e.g., home automation devices or virtual assistants, which require power-efficient classification pipelines. However, conventional classification pipelines that digitize the analog signals at Nyquist rate followed by digital feature extraction and classification are wasteful in a sense that the "feature rate" is generally much smaller than the Nyquist rate. In this paper, we propose a novel classification pipeline called analog-tofeature (A2F) conversion that directly acquires features in the analog domain using non-uniform wavelet sampling (NUWS). Our approach effectively combines Nyquist-rate sampling and digital feature extraction, which has the potential to significantly reduce the power and costs of signal classification. We demonstrate the efficacy of our approach for the detection of audio events and show that NUWS-based A2F conversion is able to outperform existing methods that use compressive sensing.

Sample-Efficient Spatio-Spectral Whitespace Detection Using Least Matching Pursuit

“…We start by generating an overcomplete NUWS base dictionary W ∈ C D×N consisting of D M different wavelets. As in the single-antenna case [30], [37], we focus on wavelet sequences with elements taken from the set {+1, 0, −1}, which has the advantage of enabling simple analog circuitry to generate the wavelet sequences. From this base dictionary, we generate an overcomplete block-diagonal base dictionary for the multi-antenna case as Q = I B ⊗ W D .…”

C-AVDI: Compressive Measurement-Based Acoustic Vehicle Detection and Identification

“…The term compressive signal processing (CSP) was first coined by the authors of [21] and refers to the process of performing detection, classification, and filtering directly on the compressive measurements obtained during the CS process without prior reconstruction. This concept has been further explored and analyzed in [22,28,29]; and while the goals and approaches differ, the fundamental idea of bypassing the reconstruction of x and instead directly leveraging y remains the same. In a similar manner, in our paper, we perform AVDI by extracting features directly from compressive measurements.…”

Identifying Unused RF Channels Using Least Matching Pursuit

“…Non-uniform wavelet sampling (NUWS) has been proposed in [16] as a flexible and hardware-friendly compressive sensing strategy for RF sensing [18] and feature extraction tasks [19]. In short, NUWS combines the advantages of nonuniform sampling [9] and random modulation [20].…”