使用和弦編碼轉換的流行音樂鋼琴樂曲自動生成

Abstract

音樂生成與影像生成及影片生成有著一些顯著的差異。首先，音樂是時間上的藝術，所以我們需要利用時序處理的方法。接著，音符不僅僅是純粹時序上的先後關係，鄰近的音群可以組成各式的音樂語法、結構，例如和弦、琶音與音階等等。本篇論文，在基於自注意力機制模型的框架下，我們探討如何生成數分鐘長的流行鋼琴音樂，我們也進一步地提出一套資料前處理的流程，藉由此流程我們可以從原始音訊轉換為音樂數位介面格式。為了分析生成結果，透過主觀的使用者問卷，我們得到許多深刻的見解，並且對模型的架構優缺點做一個通盤性的探討，進一步驗證了我們提出方法的有效性，從而了解深度學習技術的有效性與局限性。

Generating music has a few notable differences from generating images and videos. First, music is an art of the time, necessitating a temporal model. Second, musical notes are often grouped into chords, arpeggios, or melodies in polyphonic music, and therefore introducing a sequential ordering of notes into the generating model is critical. In this thesis, we investigated the framework of the Transformer model for generating minute-long pop piano music. We also proposed a data pre-processing pipeline to collect audio data and convert it to MIDI format. To evaluate the generated results, we adopted subjective user study to demonstrate the effectiveness of the proposed method.