從音樂特徵探勘到風格敘寫：生成式AI在江南音樂風格摹寫上的應用研究

The Application of Generative AI in Composing Jiangnan Style Music: From Musical Features to Stylistic Portrayal

人工智慧（AI）不僅在當今社會掀起新一波工業革命浪潮，這股科技的力量也快速延伸至音樂創作領域，目前已有許多開放式平台，如：Suno、AIVA等音樂生成工具面市，在音樂製作、編曲等領域上已有大幅運用。為了嘗試以AI技術介入傳統音樂風格的摹寫，我們以江南音樂為研究議題，首先運用「馬可夫鏈蒙地卡羅方法」來分析並統計音樂資料，利用統計方法將音樂模型的資訊特徵估計出來，再以「模擬退火演算法」技術進入江南音樂風格的摹寫。本文探討前置的基礎研究過程並運用實際生成的樂曲案例，比對人工智慧摹作樂曲的生成軌跡，藉以探討人工智慧涉入傳統音樂風格敘寫的可行性與未來性。研究發現，音樂風格可以被AI描摹並進一步生成，當風格特徵模組蒐集越完整、資料庫內容建置越豐富，AI生成的音樂風格表徵便更趨明顯。細緻的音樂特徵探勘使得AI得以在大量規則養成之下，進行特定音樂風格的敘寫；其中，樂音秩序的常態規則、樂彙組織的基本形式、韻律拍節的樣態組合，以及音階、調式、句法等關於旋律、節奏和結構要件的描述與分析設定，是人工智慧涉入音樂書寫的基礎環節，也是影響風格摹作成果的關鍵因素。

The pervasive development of Artificial Intelligence (AI) not only sparks a new wave of industrial revolution in contemporary society but also reshapes creative industries, notably in the domain of music creation. Open platforms such as Suno and AIVA are introduced and have been widely applied in areas of music production and arrangement. This study investigates the application of AI technology in the stylistic replication of traditional musical forms, specifically focusing on Jiangnan music. Initially, the Markov Chain Monte Carlo method is utilized for the statistical analysis and estimation of the formal features of music. Subsequently, the Simulated Annealing Algorithm is deployed to emulate and replicate the stylistic nuances of Jiangnan musical convention. By reviewing and applying results from actual music generation cases, this research compares the generative trajectories of AI-composed music against the source composition, thereby assessing the feasibility and future potential of AI in reproducing traditional musical styles. The findings indicate that musical styles can be effectively replicated and further developed by AI. Stylistic fidelity is shown to increase with the sophistication of style feature modules and the expansion of the underlying musical database. Detailed feature mining enables AI to narrate specific musical styles under a framework of comprehensive rules. Critical foundational elements for successful stylistic imitation include: the normative rules governing musical order, the primary structures of melodic phrases, combinations of rhythmic patterns, and the descriptive/analytical settings for melody, rhythm, and structural components (e.g., scales, modes, and syntax). These factors are identified as the basis for AI’s capacity in music composition and are pivotal to the outcomes of style imitation.