Memory-Efficient Modeling and Search Techniques for Hardware ASR Decoders

This paper gives an overview of acoustic modeling and search techniques for low-power embedded ASR decoders. Our design decisions prioritize memory bandwidth, which is the main driver in system power consumption. We evaluate three acoustic modeling approaches — Gaussian mixture model (GMM), subspace GMM (SGMM) and deep neural network (DNN) — and identify tradeoffs between memory bandwidth and recognition accuracy. We also present an HMM search scheme with WFST compression and caching, predictive beam width control, and a word lattice. Our results apply to embedded system implementations using microcontrollers, DSPs, FPGAs, or ASICs.