<![CDATA[The use of future context in acoustic modeling seems to give an impact on system performance such us Bidirectional Long Short-Term Memory (BLSTM). It has been used as an acoustic model on Speech Recognition System for Quran recitation and show better result than Hidden Markov Model - Gaussian mixture model (HMM-GMM) with average Word Error Rate (WER) value 4.6%. but, the architectural complexity of BLSTM make the latency during decoding process is high. To reduce the latency, Minimal Gated Recurrent Unit with Temporal Convolution (mGRUIPTC) acoustic model was used. Text data such as transcription, lexicon, and corpus used in training are represented at phone level to handle phone level detection. The transcription is generated using modified QScript to handle reciting rules in detail. In the test, the system can reduce decode process latency by up to 11 seconds with Phone Error Rate (PER) difference of up to 1.46% compared to BLSTM. However, our model still needs to be trained with more data to detect error recitation better.]]>
Copyrights © 2026
