使用序列模型將樂譜轉譯為吉他譜

Abstract

吉他指法譜的編排是一項專業技能，需要投入大量時間學習，甚至需要向專家請教。初學者和不熟悉吉他的音樂創作者可能因此卻步。我們希望通過自動化的指法轉譯系統，降低指法編排的門檻，讓更多人享受吉他演奏的樂趣。

因此，本研究提出了一種將樂譜(Sheet Music)作為輸入，吉他指法譜作為輸出的機器學習模型。

由於音樂是一種具有時序且前後相關的資訊，適合使用序列模型(Sequence Model)處理，我們設計了基於Transformer架構的模型進行轉譯工作。我們採用dadaGP資料集(包含26181個不同內容的GuitarPro檔案)作為訓練資料集。我們先過濾多餘的檔案，並透過PyGuitarPro套件讀取吉他指法譜資料。基於經驗和觀察，我們選擇忽略節奏資訊，並以此前提設計了音樂資訊和指法資訊的嵌入方法。此外，我們設計了資料後處理方法，檢查輸出的指法譜，若發現某個指法對應的音高不在輸入資訊中，則改為不彈奏，從而改善模型輸出表現，並成功通過GPU平行運算實現，降低了後處理的計算時間。

在實驗中，我們首先嘗試訓練出一個能夠收斂的模型，並以此為基礎，尋找表現較好的參數和架構設定。實驗結果在表現上優於近年的研究。

Arranging guitar tablature is a specialized skill that requires a significant amount of time to learn and often involves seeking advice from experts. Beginners and music creators unfamiliar with the guitar might be discouraged by these barriers. We aim to lower the barriers of tablature arrangement through an automated tablature transcription system, enabling more people to enjoy playing the guitar.

Thus, this study proposes a machine learning model that takes sheet music as input and produces guitar tablature as output.

Since music is sequential and context-dependent, it is well-suited for sequence models. We designed a model based on the Transformer architecture to handle the transcription task. We used the dadaGP dataset, which contains 26,181 unique GuitarPro files, as our training dataset. We filtered out redundant files and read the guitar tablature data using the PyGuitarPro library. Based on experience and observations, we ignored rhythmic information and designed our embeddings for musical and tablature information accordingly. Additionally, we developed a post-processing method to check the output tablature. If a note in the tablature does not correspond to a pitch in the input, we mark it as unplayed, thus improving the model's performance. This process was successfully implemented using GPU parallel computing, reducing the computation time for post-processing.

In our experiments, we first trained a convergent model and then used it as a baseline to find better parameter settings and model configurations. The results outperformed recent study in terms of performance.