Efficient PDM Microphone Array Processing
Nowadays, pulse-density modulated (PDM) digital microphones are widely used in commercial applications as they have become a popular way to deliver audio to digital processors on mobile applications. However, as state-of-the-art array processing algorithms assume that all microphone signals are available in pulse-code modulated (PCM) representation, these digital microphones require costly high-order decimation filters to translate PDM bitstreams to baseband multi-bit PCM signals. In that manner, the implementation of microphone array algorithms in embedded systems, where processing resources are critical, or in very-large-scale integration (VLSI) circuits, where power and area are critical, may become very expensive because of the use of the tens of decimation filters required to convert PDM bitstreams into PCM signals. This thesis explores and proposes resource-efficient methods to implement microphone array beamforming. For this purpose, it first reviews the state-of-the-art decimation filter design methods and proposes an algorithm to design decimation filters optimizing area and power consumption. Then it discusses the trade-offs of doing the beamforming calculations at the PDM bitstream instead of PCM signals and proposes an architecture to implement beamformers without decimation filters. Finally, it proposes an efficient approach to implement beamformer based on maximally flat (MAXFLAT) filters. As a result, a new generalized method to design decimation filters optimizing the number of additions per second is proposed, and it is shown that a beamformer implemented in the PDM domain requires fewer resources for its implementation in the time domain than other methods. It is concluded that the proposed MAXFLAT-based approach has better storage versus computation efficiency than state-of-the-art and PDM domain implementation approaches.