Deep Learning-based Syllable Recognition Framework for Korean Children

Recently, speech recognition has been very widely used in our life. With this trend, this work introduces a new recognition system, called SCKC (Syllable Classifier for Korean Children), for Korean syllable, especially spoken by children. In particular, to design the recognition system, we first collected 10,000 Korean syllable data from approximately 150 children. Based on observations on the dataset, we optimize each procedure of SCKC, including pre-processing, feature extraction, and deep CNN model construction. A prototype implementation of SCKC provides a reasonable degree of accuracy (85.4% on average) in syllable recognition.