Speaker
Description
(Overview and Significance of the Study)
Infant language development[1] accelerates rapidly between the ages of 5 and 7. The ability to accurately communicate one's intentions is a necessary component of interpersonal relationships and social interaction. According to KOSIS[2], in 2020, approximately 15.4% of children aged 0-9 years had a pronunciation disorder (articulation disorder). Articulation disorder is the inability to produce speech sounds correctly due to structural or functional problems with the articulatory organs such as the lips, tongue, teeth, and soft palate. One of the ways to improve articulation accuracy in these children is to provide repetitive experiences through play. Different experiences require different content to be developed.
(Main features)
The game is a word puzzle consisting of prearranged words. The player selects a puzzle, and when the player correctly reads the chosen word, the puzzle disappears, creating a path. Water flows along this path. Throughout the puzzles, there are obstacles like lava that hinder the flow of water. If the player successfully guides the water to its destination, the stage is completed. There are a total of 5 levels, each containing 10 stages. The difficulty increases with longer word lengths and deeper puzzle depths as the levels progress.
(Service operation process)
The user inputs the pronunciation of a word in a game. The inputted audio file is then converted to text through Naver's Clova API. The converted text is sent to a server created with Django. Afterwards, an artificial intelligence model is used to compare the accuracy of the language's consonants with the accuracy of the word presented as a problem. If the accuracy exceeds the accuracy required for each level, it becomes the correct answer, and if not, it is recorded in the incorrect problem in the database. The correct pronunciation of the answer is provided to the user using the Clova API.
(Data collection)
Data collection involves collecting words for the AI model to learn and solve problems. Words for the AI model to learn will be collected from news data, daily conversation data, Wikipedia data, and elementary school Korean language textbooks. Problems to be provided during the game will be classified and collected from the words in elementary school Korean language textbooks that are suitable for young children [3].
(Data processing)
Data processing involves using the Clova API to convert speech to text, separating vowels and consonants using the Jamo library, and storing the data in a relational database along with accuracy determined by the AI model.
(Applications of AI models)
The AI model we will use is word-to-vector [4], which represents words as vectors and expresses the relationships between vectors as words with similar vowels and consonants instead of semantic relationships. We will use Tensorflow as the framework and the trained model will receive text data, compare it with the correct label, and determine accuracy.
(Application strategy)
Our app's strength is visual enjoyment. Building upon this, we can not only develop an app for treating children's Korean pronunciation but also convert the learning data into English, which can be utilized for an app to study English pronunciation.
References
[1] Yoo, J. K., Lee, K. O., & Lee, K. M. (2012). Speech Database for 3-5 years old Korean Children. The Journal of the Korea Contents Association, 12(4), 52-59.
[2] KOSIS (Prevalence of Speech Impairment at Home by Age)
[3] Mi-Sun Yoon., & Seunghwan Lee (1998). A Comparative Study on the Measures of Intelligibility and Percentages of Consonants Correct Between Phonologically Disordered and Normal Children. Communication Sciences and Disorders, 3(1), 50-68.
[4] Tomas Mikolov, Kai Chen, Greg Corrado, Jeffrey Dean(2013) , Efficient Estimation of Word Representations in Vector Space, Conference on Empirical Methods in Natural Language Processing (EMNLP)
| Keywords | infant pronunciation disorder, a digital remedy, a medical game, contents for children |
|---|
