Scholarlink

<p>Deep learning, particularly Neural Networks, has demonstrated strong capabilities in analyzing, detecting, and recognizing human activities, with widespread applications across various industries, including entertainment. Leveraging deep learning to predict song names from short video clips featuring dance movements and “E”-monosyllable sounds opens up the possibility of developing automatic trivia games that enhance user experience and offer commercial value. Despite its potential, this area remains underexplored, especially in the context of classifying 90s Thai songs. To address this gap, we curated a custom dataset of 750 video clips from 15 well-known 90s Thai songs, chosen for their distinctive dance styles and continued popularity. In this study, we developed three models: an audio-unimodal model using long short-term memory (LSTM), a visual-unimodal model using convolutional neural network (CNN), and an audio-visual multimodal model that integrates both modalities. Experimental results show that the visual-unimodal model achieved the highest testing accuracy at 91.33%, while the audio-unimodal model performed significantly lower at 47.33%. The multimodal model performed comparably to the visual model, with an accuracy of 91.11%. These findings suggest that visual-based approaches are effective and practical for recognizing dance movements associated with 90s Thai songs. Moreover, the proposed visual model has the potential to generalize to international songs that exhibit distinctive dance characteristics.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

90s Thai music classification using audio-visual multimodal model

  • 期刊论文

  • 发表日期：

    2025年6月1日

查看全文
  • Kanis Charntaweekhun,
  • Duangkamon Ketchanchai,
  • Jittikan Narapan,
  • Penprapa Wutthijuk,
  • Sasithorn Sirintrawutthiwong,
  • Thitirat Siriborvornratanakul

摘要

Deep learning, particularly Neural Networks, has demonstrated strong capabilities in analyzing, detecting, and recognizing human activities, with widespread applications across various industries, including entertainment. Leveraging deep learning to predict song names from short video clips featuring dance movements and “E”-monosyllable sounds opens up the possibility of developing automatic trivia games that enhance user experience and offer commercial value. Despite its potential, this area remains underexplored, especially in the context of classifying 90s Thai songs. To address this gap, we curated a custom dataset of 750 video clips from 15 well-known 90s Thai songs, chosen for their distinctive dance styles and continued popularity. In this study, we developed three models: an audio-unimodal model using long short-term memory (LSTM), a visual-unimodal model using convolutional neural network (CNN), and an audio-visual multimodal model that integrates both modalities. Experimental results show that the visual-unimodal model achieved the highest testing accuracy at 91.33%, while the audio-unimodal model performed significantly lower at 47.33%. The multimodal model performed comparably to the visual model, with an accuracy of 91.11%. These findings suggest that visual-based approaches are effective and practical for recognizing dance movements associated with 90s Thai songs. Moreover, the proposed visual model has the potential to generalize to international songs that exhibit distinctive dance characteristics.