許多的大型音樂網站為了方便使用者檢索及查詢都會提供音樂分類目錄功能,但使用手工分類不但高成本且沒有效率。另外對於沒有樂理基礎的人要為主旋律搭配和弦非常不容易。因此本研究的目的是為單純的主旋律,判斷最適合的曲風自動配置和弦伴奏,並可對已經擁有和弦的音樂,辨別其曲風作音樂分類。 本研究與先前MIDI音樂分類研究不同,不只使用音高、音長…等特徵。更將擷取的資訊再進一步辨識出每小節的主旋律及和弦特徵,並且考慮和弦演奏方式和避免非和弦組成音干擾的權重計算方式;然後建立相連和弦的Hidden Markov Model,在使用者利用系統輸入主旋律後,以Viterbi Search找到最適合的和弦序列自動配置和弦伴奏,也可以利用訓練出的HMM對未知曲風的音樂做分類。 分類實驗所使用的訓練資料包含古典、鄉村、搖滾及爵士樂四類,共365首音樂;首先使用G. Frederico(2004)研究中的測試資料,共15首歌曲分為三類,平均正確率為93.3%,較G. Frederico實驗進步11.5%。接著使用十次交叉驗證分三類正確率為61.51%。最後用大資料集,共120首歌曲分為四類,實驗結果顯示如果只分古典、鄉村、爵士三類平均正確率可達62.16%。 和弦伴奏成效評估方式,是將本研究和Band in a Box(BIAB)音樂軟體自動和弦配置音樂匯出,由受訪者直接聆聽評分。結果顯示在9首音樂中,其中4首評分高於BIAB軟體自動伴奏結果,其餘5首評分差距也在1分以內。評分滿分為10分,BIAB總平均為6.2分,本研究為5.9分僅差距0.3分。
With the development of music software, composition is no more the proprietary work to musicians. However, it is difficult for people without knowledge of music theory to make accompaniment. Besides, many music websites provide the music style indexes for users to quickly retrieval and query. But it is also costly and inefficiently to classify music manually. The goal of this research is modeling music styles that can do the melody style classification and then make the most appropriate accompaniment automatically. Different from the past researches, our study extracts chord features in the ways that consider the performance of chords and weight the possible chords to avoid the disturbance of non-chord noise and then to establish the Hidden Markov Models (HMMs). After user input melody, the system can find the most suitable sequence of the chord accompaniment by Viterbi search on HMMs and can also be used to classify those accompanied music by HMMs.