PeakEngine: A Deterministic On-the-Fly Pruning Neural Network Accelerator for Hearing Instruments

Recurrent neural networks (RNNs) are well-suited for sequential tasks such as speech enhancement (SE). However, their performance comes with high-computational complexity and latency. This impedes their deployment to battery-powered and resource-constrained hearing instruments (HIs) that need to operate for 16–18 h daily at only a few milliwatts (mW). In this article, we introduce PeakEngine, a configurable ASIC accelerator that decreases the amount of computation and memory accesses, and thus latency, in a gated recurrent unit (GRU) by means of adaptive inference. The reduction is achieved by on-the-fly pruning that selects the top
$K$
elements based on magnitudes of delta changes across timesteps from both input and hidden state sequences. Since
$K$
is constant, it results in a deterministic execution time. PeakEngine is synthesized in a 22-nm CMOS process, and the simulations show that it dissipates
$11.83 ~\mu \text{J}$
per inference for the baseline (unpruned) network and only 4.14–
$5.04 ~\mu \text{J}$
for the pruned networks, with maximum acceptable degradation to no degradation in the improvement in audio quality and intelligibility. Moreover, the inference is on average sped up 2.2–
$2.97\times $
, hence meeting the real-time requirements imposed by a HI application. To the best of our knowledge, PeakEngine is the first ASIC accelerator for deterministic and dynamic pruning in RNNs targeting HIs and SE.