Development of Output Correction Methodology for Long Short Term Memory-Based Speech Recognition

Abstract

This paper presents a correction methodology for Long Short Term Memory (LSTM) based speech recognition. A strategy that validates with a reference database was developed for LSTM. It is conceptually simple but requires a large keyword database to match test templates. The correction method is based on the “most matching method” that is finding the word in which the system output is closest among the “Referenced Template Database”. Each LSTM model recognition output was corrected with the proposed new concept. Thus, system recognition performance was improved by correcting faulty outputs. The effectiveness, efficiency, and contribution of this approach to system performance were demonstrated by experiments. Tests carried out using different speech-text datasets and LSTM models yielded an average performance increase of 2.25%. With some advanced models, this ratio rises to 3.84%.