Profiling Large-Vocabulary Continuous Speech Recognition on Embedded Devices: A Hardware Resource Sensitivity Analysis

摘要

When deployed in embedded systems, speech recognizers arenecessarily reduced from large-vocabulary continuous speechrecognizers (LVCSR) found on desktops or servers to fit thelimited hardware. However, embedded hardware continues toevolve in capability; today's smartphones are vastly morepowerful than their recent ancestors. This begets a newquestion: which hardware features not currently found ontoday's embedded platforms, but potentially add-ons totomorrow's devices, are most likely to improve recognitionperformance? Said differently – what is the sensitivity of therecognizer to fine-grain details of the embedded hardwareresources? To answer this question rigorously andquantitatively, we offer results from a detailed study ofLVCSR performance as a function of microarchitectureoptions on an embedded ARM11 and an enterprise-class IntelCore2Duo. We estimate speed and energy consumption, andshow, feature by feature, how hardware resources impactrecognizer performance.