Scholarlink

This research proposes a novel approach that leverages the transformer architecture to create a dynamic MIDI composition that responds to events in near real-time. The model leverages Rotary Positional Embeddings (RoPE) to allow infinite generation for background music applications. Custom rule-based logit filters are also provided to improve generation quality. Byte-Pair Encoding is used to improve the attention context strength. This approach has been implemented and tested with Unreal Engine integration, allowing for the synthesis and mixing of sounds within Metasound.

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

Transformer Based Adaptive Music Generation for Video Games

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Thomas Dallard,
  • Akinori Ito

摘要

This research proposes a novel approach that leverages the transformer architecture to create a dynamic MIDI composition that responds to events in near real-time. The model leverages Rotary Positional Embeddings (RoPE) to allow infinite generation for background music applications. Custom rule-based logit filters are also provided to improve generation quality. Byte-Pair Encoding is used to improve the attention context strength. This approach has been implemented and tested with Unreal Engine integration, allowing for the synthesis and mixing of sounds within Metasound.