Scholarlink

<p>The Internet of Musical Things (IoMusT) extends the Internet of Things paradigm to networked musical devices, browser-based interfaces, and cloud-connected creative services. Yet non-professional creators still face two persistent challenges: choosing an arrangement style that matches their musical intent and transforming a sparse lead melody into a coherent multi-track production. This paper presents an AI-enabled IoMusT platform that integrates connected MIDI input, context-aware style recommendation, automatic orchestration, and collaborative web-based editing. The recommendation module aligns piano-roll melody features and textual intent labels with a style library through contrastive learning. The orchestration module employs a hierarchical Transformer with explicit chord-aware conditioning to generate bass, drums, harmonic pads, and supporting tracks. At the system level, the platform combines browser-side MIDI capture, WebSocket-based session synchronization, a FastAPI gateway, and GPU-backed inference services for interactive creation. On a held-out test set, the recommendation model achieves Precision@5 of 0.83 and NDCG@10 of 0.88, while the orchestration model reaches chord accuracy of 87.6% and rhythm histogram overlap of 0.89. Improvements over the strongest baselines are supported by significance testing, independent validation, and multi-run variance reporting where applicable. In a double-blind listening study with 15 professional musicians, 73.1% of generated accompaniments were rated as plausibly human-composed; given the modest panel size we treat this as preliminary evidence rather than a claim of human-level composition. Deployment measurements from 200 user sessions show a median recommendation latency of 340 ms and a median orchestration latency of 2.1s, complemented by a concurrency and low-resource scalability analysis, supporting near-real-time IoMusT interaction. We emphasise that the methodological novelty of this work lies in two model components—a creator-facing contrastive arrangement recommender and an explicit harmonic-consistency module for orchestration—whereas the connected web stack is presented as engineering integration rather than an algorithmic contribution.</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

An AI-enabled Internet of Musical Things platform for interactive arrangement recommendation and automatic orchestration

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月10日

查看全文
  • Hengyang Xu

摘要

The Internet of Musical Things (IoMusT) extends the Internet of Things paradigm to networked musical devices, browser-based interfaces, and cloud-connected creative services. Yet non-professional creators still face two persistent challenges: choosing an arrangement style that matches their musical intent and transforming a sparse lead melody into a coherent multi-track production. This paper presents an AI-enabled IoMusT platform that integrates connected MIDI input, context-aware style recommendation, automatic orchestration, and collaborative web-based editing. The recommendation module aligns piano-roll melody features and textual intent labels with a style library through contrastive learning. The orchestration module employs a hierarchical Transformer with explicit chord-aware conditioning to generate bass, drums, harmonic pads, and supporting tracks. At the system level, the platform combines browser-side MIDI capture, WebSocket-based session synchronization, a FastAPI gateway, and GPU-backed inference services for interactive creation. On a held-out test set, the recommendation model achieves Precision@5 of 0.83 and NDCG@10 of 0.88, while the orchestration model reaches chord accuracy of 87.6% and rhythm histogram overlap of 0.89. Improvements over the strongest baselines are supported by significance testing, independent validation, and multi-run variance reporting where applicable. In a double-blind listening study with 15 professional musicians, 73.1% of generated accompaniments were rated as plausibly human-composed; given the modest panel size we treat this as preliminary evidence rather than a claim of human-level composition. Deployment measurements from 200 user sessions show a median recommendation latency of 340 ms and a median orchestration latency of 2.1s, complemented by a concurrency and low-resource scalability analysis, supporting near-real-time IoMusT interaction. We emphasise that the methodological novelty of this work lies in two model components—a creator-facing contrastive arrangement recommender and an explicit harmonic-consistency module for orchestration—whereas the connected web stack is presented as engineering integration rather than an algorithmic contribution.