An Empirical Study on Speech Recognition Performance in Low-Resource Radio Environments
Yunzhao Lu, Sam W. O. Lam, Justin Y. Y. Lee, Jie Ji, Asa M. C. Chan · 2023
This work presents the development of a robust speech recognition engine for a railway communication system using radio signals with a sampling frequency of no more than 8000 Hz, which can perform robust speech transcription in noisy environments. A domain-specific language model is constructed from limited training data and domain-specific corpus based on finetuned weights. And this corpus is created from a template of spreadsheet containing a Jargon table of common terms and a Corpus of common sentences tuned as conversation aligned to radio communication protocols. For evaluation, it is reported the model’s performance is robust comparing to human accuracy on a portion of trained data; while for un-trained data, the model out-performs human by 83.8% vs. 82% in accuracy. The design architecture of the speech recognition engine provides a robust and reliable solution for low-resource radio signal communication and can even out-perform human on poor quality data.