Real-Time and Low-Power Streaming Source Separation Using Markov Random Field

Machine learning (ML) has revolutionized a wide range of recognition tasks, ranging from text analysis to speech to vision, most notably in cloud deployments. However, mobile deployment of these ideas involves a very different category of design problems. In this article, we develop a hardware architecture for a sound source separation task, intended for deployment on a mobile phone. We focus on a novel Markov random field (MRF) sound source separation algorithm that uses expectation-maximization and Gibbs sampling to learn MRF parameters on the fly and infer the best separation of sources. The intrinsically iterative algorithm suggests challenges for both speed and power. A real-time streaming FPGA implementation runs at 150MHz with 207KB RAM, achieves a speed-up of 22x over a software reference, performs with an SDR of up to 7.021dB with 1.601ms latency, and exhibits excellent perceived audio quality. A 45nm CMOS ASIC virtual prototype simulated at 20MHz shows that this architecture is small (<10 million gates) and consumes only 70mW, which is less than 2% of the power of an ARM Cortex-A9 software version. To the best of our knowledge, this is the first Gibbs sampling inference accelerator designed in conventional FPGA/ASIC technology that targets a realistic mobile perceptual application.