SPEECH RECOGNITION USING JULIUS BASED ON INDONESIAN LANGUAGE AND ITS IMPLEMENTATION ON APPLICATION SOFTWARE
The author’s final project is about developing speech recognition to recognize sound or speech that using Indonesian language. Speech Recognition Engines (SREs) which have already developed in this world, commonly use English language as language that will be recognized. <br /> <br /&g...
Saved in:
Main Author: | |
---|---|
Format: | Final Project |
Language: | Indonesia |
Online Access: | https://digilib.itb.ac.id/gdl/view/15728 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Institut Teknologi Bandung |
Language: | Indonesia |
Summary: | The author’s final project is about developing speech recognition to recognize sound or speech that using Indonesian language. Speech Recognition Engines (SREs) which have already developed in this world, commonly use English language as language that will be recognized. <br />
<br />
<br />
So that Speech Recognition Engine could recognize Indonesian speech and language, then we need to make phoneme list, dictionary, acoustic model, and grammar/language model manually. Phoneme list must be appropriate with phonemes that used in Indonesian language. Dictionary was made manually and consist of words that will be used on application. Acoustic model was made from several sound samples that created by author. Grammar/Language model that we use must be appropriate with Julius SRE format and consist of words that will be used on application. According to the test result, test on rawfile with using HTK had sentence correct percentage about 6,38%, word correct percentage about 98,67%, and accuracy <br />
<br />
<br />
percentage about 79,20%. Then test on rawfile with using Julius had sentence correct percentage about 93,62%, word correct percentage about 98,23%, and accuracy percentage about 98,23%. To make the result better, we modified the rawfile and created new acoustic model, then the new acoustic model was tested using HTK and Julius, then gave result correct percentage about 100%. Next test was using input from microphone. Speech sound was recognized with sentence correct percentage about 94,44% and word correct percentage about 99,64%. Because of this high correct percentage, wrong result that occur on application is infrequently. In this case, wrong recognition usually occur because of OOV (Out Of Vocabulary). |
---|