An article with a main author from TCS (Tata). Talks about trying to reduce neural network model sizes while retaining high accuracy for edge devices.

A Low footprint Automatic Speech Recognition System For Resource Constrained Edge Devices

ABSTRACTDeep Learning (DL) has been instrumental in pushing artificialintelligence (AI)/ machine learning (ML) algorithms to edge of thenetwork. It allows building AI/ML algorithms for computer vision,speech processing, and other timeseries analytics tasks with limiteddomain knowledge. As there is no mechanism to control the repre-sentations learned from a large dataset, it becomes hard to predictwhether a very small DL model can learn the proper dependenciesneeded for a particular problem at hand.With speech recognition capability becoming important in sev-eral Internet of Things (IoT) devices, we propose an explainableAI-based methodology to build small DL models for speech recog-nition by controlling the representations learned by a model undera hard size constraint.We enhance the architecture of a state of the art sequence trans-duction model to allow the tuning of accuracy vs. model size trade-off. Using these techniques we achieve a reduction in model sizeand latency by a factor of 10 and 6 respectively, with only 4losscompared to the embedded implementation of a well known ASR.

https://www.google.com/url?sa=t&source=web&rct=j&url=https://dl.acm.org/doi/abs/10.1145/3417313.3429385&ved=2ahUKEwic35iM2bHtAhV2zTgGHW_SDoc4ChAWMAB6BAgDEAE&usg=AOvVaw1nVN2sP1_CxRxkUzF78dYl

DYOR