DEVELOPMENT OF ACOUSTIC MODEL USING DNN BASED ON CHAIN MODEL IN SPEECH RECOGNITION SYSTEM (CASE STUDY: DENTAL CONVERSATION)

• Main Author: Yora Islami, Dinda
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/58052
• Institution: Institut Teknologi Bandung

The problem of incomplete dentistry medical records due to time constraints can be solved by developing a speech recognition system in dentistry domain to write dentist medical records. The speech recognition system in dentistry domain has been developed by researchers Islami & Lestari (2020) using Convolutional Neural Network (CNN) technique to develop acoustic models and n-gram technique to develop language models. System performance was measured by word error rate (WER) and the WER value was 14.47%. This performance can be improved by developing acoustic models and/or language models with other techniques and by handling out of vocabulary (OOV) problems. In this study an acoustic model was developed with other techniques that can improve the performance of the speech recognition system in dentistry domain. In several researchers the development of an acoustic model using the Time Delay Neural Network (TDNN) or Long Short-Term Memory (LSTM) technique resulted in better performance than the CNN technique. Then the development of an acoustic model with a chain model can improve the performance of the speech recognition system. So, in this study an acoustic model was developed using CNN, TDNN, LSTM and the application of the chain model to the three techniques. The language model in this study uses the n-gram technique, and the handling of OOV problems is solved by adding the dental domain text corpus. The results of this study showed that the treatment of OOV with the addition of the dental domain text corpus did not significantly reduce the OOV value. The development of acoustic models using TDNN, and LSTM techniques has not been able to outperform the baseline (CNN) technique. The development of an acoustic model with a chain model on the TDNN technique results in good performance compared to other techniques and baseline. In the TDNN chain, the WER value decreased by 3.6% from the baseline, with a WER value of 10.85%.