基於歌詞情緒的華語流行歌曲自動作曲系統

Abstract

本論文目標為實作一個能依據使用者所輸入的歌詞，分析歌詞情緒並按使用者所指定的情緒進行自動作曲的系統。歌詞分析會先對使用者輸入的歌詞進行自動斷詞，再經由情感詞彙計算出各種特徵值，最後以專家標定的歌詞情緒訓練資料進行SVM分類。自動作曲則會依照使用者輸入的歌詞做二維情緒平面正向度的辨識，提供界面推薦使用者選擇情緒，再依其情緒參數產生相對應的和弦進行、歌曲速度、音域範圍，由上述條件透過以訓練資料產生的一階馬可夫鏈表，進行自動作曲並產生Midi檔。 此研究依目前已發行的歌詞進行自動作曲，經回饋問卷得知，滿分5分，情緒歌曲相關程度為3.33分。本系統特以Swift程式語言為開發語言，在iOS系統實作，以期未來能更方便的被使用及擴充。本研究的主要用途有二，第一是希望能降低音樂創作的門檻，讓文字創作者也有作曲的機會。第二是希望能提供創作者更多的選擇性和靈感，提升創作的效率。

The aim of this thesis is to build a system which can compose automatically based on the lyrics input by users, the sentiments of lyrics the system analyze, and the emotions the users assign. While analyzing the lyrics, the system would segment the sentences automatically first, and then calculate all kinds of feature values via emotion vocabulary, and finally conduct SVM assortments based on the lyrics emotion dataset set by experts. Automated Composition would proceed with the valence according to the emotion of given lyrics, and provide interfaces for users to choose emotions from. Then, in accordance with the relative chords, song tempo and note range resulted from the emotion feature, the system compose automatically and produce MIDI via the 1-st order Markov Chain brought out from training data。 This study composes songs with the lyrics which have been released, and according to the survey feedback, satisfaction with emotion songs is 3.33. This system uses Swift as the developing language and is operated by iOS in order to be used and extended more conveniently in the future. The purposes of this study are: 1. to hopefully lower the threshold for music-making and offer the lyric writers opportunities to compose; 2. to provide composers with more choices and inspiration to enhance the composing efficiency.