從音訊到樂譜與音色：探討以吉他為核心的音樂資訊 檢索中的表示法與轉換

Abstract

電吉他是現代音樂的重要元素，其與原聲吉他及鋼琴的最大區別，在於其音色高度依賴經由擴大器與效果器處理後所產生的各種變化。然而，相較於以鋼琴為主的研究，吉他導向的音樂資訊檢索（Music Information Retrieval, MIR）發展相對落後，主因包括資料集稀缺（受限於版權與蒐集困難）以及效果處理音訊所需的特殊表示需求。本論文致力於推進吉他導向的 MIR，提出兩項核心貢獻：首先，建立並釋出兩個新資料集──Electric Guitar Database（EGDB）及其擴充版本 EGDB-PG；其次，針對兩項關鍵表示轉換任務設計創新深度學習演算法：（1）透過電吉他自動轉譜，實現由效果處理音訊轉換為符號樂譜的 audio-to-score 轉換；（2）透過電吉他音箱音色建模，實現由乾淨音色轉換為帶有效果的音色的 clean-to-wet 轉換，並探索無監督與零樣本等深度學習場景以重建多樣音色效果。透過實驗驗證，本研究有效提升吉他音訊、樂譜與音色的表示能力，有效解決電吉他效果處理所帶來的挑戰，為更穩健的基於深度學習的吉他音樂分析與吉他音箱模擬奠定基礎。

Electric guitars, central to modern music, are distinguished from acoustic guitars and pianos by their reliance on effects processing through amplifiers and pedals, which introduces complex tonal variations. However, guitar-oriented music information retrieval (MIR) lags behind piano-oriented research due to scarce datasets, constrained by copyright and collection challenges, and the unique representational demands of effect-laden audio. This thesis advances guitar-oriented MIR by curating novel datasets, the Electric Guitar Database (EGDB) and its expanded version (EGDB-PG), and proposing original deep learning algorithms for two key representation transformations: (1) audio-to-score transformation via electric guitar transcription, leveraging EGDB and EGDB-PG to map effect-processed audio to symbolic notation; and (2) clean-to-wet audio transformation through electric guitar amplifier tone modeling, exploring unsupervised and zero-shot paradigms to replicate diverse effects. Validated through empirical evaluations, these contributions enhance audio, score, and tone representations in guitar-oriented MIR, addressing the distinct challenges posed by electric guitar effects and paving the way for robust deep-learning based electric guitar music analysis and effect modeling systems.