Search

The preservation of traditional dances as an important part of world cultural heritage can be done by recording. While it is convenience to record the dances using video, this medium has limited capability in the reconstruction. On the other hand, recording using a motion capture device gives us the ability to replay them and add alterations in a creative process. In this paper, we propose a method to train traditional dance moves in a generative model using Long Short-Term Memory (LSTM). We use a traditional dance from East Java, Indonesia, that is called Remo Dance as the training data. The dance is recorded with a motion capture device and each basic move is trained into the model. In the sampling process, the trained model reiterates its memory into an unlimited length of dance animation. The generated dance animation has imperfection relative to the training data. This discrepancy gives the intended variations. We use visual assessments, dynamic time warping curves, and a subset of parameters from Laban motion analysis to evaluate the variations. These evaluations show how the variations behave and in what pattern they occur. In general, those variations give slight alterations to the motions that add human-like imperfection and give opportunities for animators and choreographers alike to explore new dances creations [ABSTRACT FROM AUTHOR]

The preservation of traditional dances as an important part of world cultural heritage can be done by recording. While it is convenience to record the dances using video, this medium has limited capability in the reconstruction. On the other hand, recording using a motion capture device gives us the ability to replay them and add alterations in a creative process. In this paper, we propose a method to train traditional dance moves in a generative model using Long Short-Term Memory (LSTM). We use a traditional dance from East Java, Indonesia, that is called Remo Dance as the training data. The dance is recorded with a motion capture device and each basic move is trained into the model. In the sampling process, the trained model reiterates its memory into an unlimited length of dance animation. The generated dance animation has imperfection relative to the training data. This discrepancy gives the intended variations. We use visual assessments, dynamic time warping curves, and a subset of parameters from Laban motion analysis to evaluate the variations. These evaluations show how the variations behave and in what pattern they occur. In general, those variations give slight alterations to the motions that add human-like imperfection and give opportunities for animators and choreographers alike to explore new dances creations. [ABSTRACT FROM AUTHOR]

In a communication using texts input, a phoneme must be mapped to viseme (visual phoneme). Phoneme is the smallest element of a language that can differentiate a meaning. Viseme is derived from a group of phonemes having similar visual appearances, the equivalent unit in the visual domain that models a speech recognition system audio-visually. This paper proposes a classification of visemes in Indonesian Language and establishes phoneme-viseme mapping for Indonesian Language. In Indonesian words, there are lots of absorbed words from other languages. They come from vernacular or foreign languages, such as Arabic and English. From 49 phonemes, 12 Indonesian visemes have been produced, including silent. Viseme classes are defined through linguistic knowledge and grouped by the same visual appearance, and then validated through a survey. The approach used in Indonesian phoneme-to-viseme mapping is based on linguistic data. This is the first research in Indonesian phoneme-to-viseme mapping, which is expected to grow into a reference for further development in human language technology. [ABSTRACT FROM AUTHOR]