利用稀疏表示法之自動鋼琴轉譜

Abstract

音高資訊和其他中階特徵被視為跨越語意隔閡的關鍵資訊，因其與人類感知十分密切相關。本篇論文提出了一個極為有效的解決方法來估計鋼琴音樂中的音高資訊，並將此資訊轉為樂譜。實驗結果證實，我們的系統其轉譜準確率目前居世界領先地位。我們將轉譜問題以頻譜的稀疏表示法來解決。我們所提出的系統首先找到有可能出現的音高，再透過解l1最小化問題及隱馬可夫模型，抽取出鋼琴音樂中每個音符的音高及持續時間，完成轉譜。

Like rhythm and timbre, pitch as a mid-level music feature holds the promise of bridging the well-known semantic gap between low-level features and high-level semantics of music. Pitch estimation is an important first step towards this ultimate goal. In this thesis, we target the extraction of multiple pitch contours from piano music signals. Specifically, the pitch estimation is formulated as a sparse representation problem, in which the feature vector of a piano music segment (or frame) is represented as a linear combination of the feature vectors of individual piano notes. The note candidates of the input piano music segment are determined according to the harmonic structure of piano sounds. Then, the sparse representation problem is solved by l1-regularized minimization. A post-processing method based on hidden Markov models (HMMs) is applied to the resulting sparse vector for accuracy refinement. The system performance is evaluated using l1 classical music recordings of a real piano. The results show that the proposed system outperforms three state-of-the-art systems.