Low-Latency Kazakh Speech Recognition on Edge Devices

Abstract

We report a quantised streaming recogniser for Kazakh that runs on a 6 W ARM board without a network connection. Eight-bit weight quantisation and a 320 ms chunked encoder bring median end-to-end latency to 180 ms at a word error rate of 12.4 per cent on the KSC2 test set, against 11.8 per cent for the unquantised model on a server GPU. The result matters for classroom and field devices in places where connectivity cannot be assumed. Model and evaluation scripts are available on request from the authors.