Low-Latency Kazakh Speech Recognition on Edge Devices

Assel Nurgaliyeva; Timur Iskakov

Low-Latency Kazakh Speech Recognition on Edge Devices

Авторлар

Assel Nurgaliyeva Qazaq AI Research University, Astana, Kazakhstan Автор https://orcid.org/0000-0002-6640-1750 ##orcid.unauthenticated##
Timur Iskakov Qazaq AI Research University, Astana, Kazakhstan Автор https://orcid.org/0000-0003-8812-064X ##orcid.unauthenticated##

Кілт сөздер:

speech recognition, Kazakh language, edge computing, quantisation, latency

Аңдатпа

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.