以和諧度優化指標設計的自動譜和弦系統

Harmony-Optimized Design for Automatic Chord Generation System

洪啓恩(Chi-En Hung)

指導教授 ： 張文輝

國立交通大學/電機學院/電信工程系所/碩士(2019年)

摘要

針對特定旋律進行伴奏和弦的編曲是業餘創作者的一項挑戰，不僅需要樂器演奏的經驗，也必須具備音樂和聲學的專業知識。本論文旨在針對業餘音樂人的編曲創作，開發一個能為輸入主旋律搭配伴奏音樂的自動譜和弦系統。有別於前人研究聚焦於原創曲與自動生成和弦的正確率，我們引入能客觀評量其和聲相似度的調性音高步距，並開發以和諧度優化為設計目標的和弦生成系統。首要之務是分析華語流行歌曲的旋律與和弦之和聲相關性，並據以建構一個能描述其和弦行進的隱藏式馬可夫模型。為了提升和弦生成系統的整體效能，我們針對隱藏式馬可夫模型的觀察機率建模程序，採用了神經網路以及基於受限波茲曼機的預訓練演算法。此外，我們也擴大定義模型的隱藏狀態，使其兼顧和弦狀態與調性音高空間狀態的行進規律，並在維特比解碼演算法中引入調性音高步距作為和諧度優化的設計指標。針對華語流行音樂的實驗結果顯示，依和諧度優化指標設計的自動譜和弦系統，能有效提升伴奏音樂的和聲相似度。在人耳聆聽主觀測試上，自動生成的伴奏音樂也廣為一般大眾喜好。

關鍵字

自動譜和弦 ； 調性音高步距 ； 隱藏式馬可夫模型 ； 和諧度優化 ； 神經網路 ； 受限波茲曼機 ； 調性音高空間 ； 維特比演算法

並列摘要

It is a challenge for amateur composers to craft proper chord arrangements to specific melodies, due to lack of the training in harmony and instrument playing. This study aims to develop an automatic chord generation system which helps amateurs arrange chord to match the input melody. Unlike the previous works focusing on the matching accuracy between the original and the generated chord sequences, our goal is to optimize their harmony similarity with the aid of an objective measure called the tonal pitch step distance (TPSD). The system design is based on a hidden Markov model (HMM), in which the hidden state represents the chord and the observation vector is derived from the melody. Several methods have been proposed to further enhance the system performance. First, the observation probability model of HMM is constructed through the use of a neural network and a pre-training algorithm, based on the restricted Boltzmann machine. Secondly, the hidden state is expanded to include both the chord state and tonal pitch space state. Finally, a TPSD-modified Viterbi algorithm is proposed for decoding of the hidden state sequence. Experiments on Chinese pop songs indicate that the proposed methods significantly improve the harmony of a HMM-based chord generation system. In addition, the subjective listening test also demonstrates that most people prefer the accompaniment music generated by the proposed system.

並列關鍵字

automatic chord generation ； tonal pitch step distance ； hidden Markov model ； harmony-optimized ； neural network ； restricted Boltzmann machine ； tonal pitch space ； Viterbi algorithm

參考文獻

[1] I. Simon, D. Morris, and S. Basu, “MySong: automatic accompanimentgeneration for vocal melodies,” Proceedings of ACM CHI 2008 Conferenceon Human Factors in Computing Systems, vol. 1, pp. 725–734,2008.

[2] D. Morris, I. Simon, and S. Basu, “Exposing parameters of a traineddynamic model for interactive music creation,” Proceedings of AAAI 2008, no.2003, pp. 784–791, 2008.

[3] Gunhild Elisabeth Berget,“Using Hidden Markov Models for MusicalChord Prediction”, NTNU, 2017.

[4] C.-H. Chuan and E. Chew, “A Hybrid System for Automatic Generationof Style-Specific Accompaniment,”Proceedings of the 4th InternationalJoint Workshop on Computational Creativity, pp. 57–64, 2007.

[5] G. Hinton, L. Deng, D. Yu, G. Dahl, A. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke, P. Nguyen,T. Sainath, and B. Kingsbury,“Deep neural networks for acoustic modeling in speech recognition,” IEEESignal Processing Magazine, 2012.