樂器音源訊號之分析及Wavetable之合成方法

Abstract

在電腦音樂中，音樂的合成有效地降低了記憶容量。以頻譜為主的分 析電腦音樂方式是常用的方式，大部分樂器的頻譜都是隨時變的。原始音 源的時變頻譜分析是合成的主要步驟。透過瞬時傅立葉分析，吾人能夠穫 得樂音的完整隨時變頻譜。 以效能觀點而言，Wavetable 合成模式是 常用的合成方法。在尋找基底時，我們使用四種方法（ 分群法、主軸分 析法、分群追加法、重心映成法 ）來探索隨時變頻譜的特徵。對於各種 樂器而言，超過百分之五十的資料量被降低了，而所合成的樂音和原來的 樂音之間是吾人無法分辨出來的。對於同一樂器中不同音調的樂音，我們 試圖尋找一組共通性的︱avetable 來合成這些樂音。 在本論文□， 我們以績效和複雜度來比較此四種方法。在找尋基底時，分群法是其中最 容易的方式，並且有很好的效能。但是這種方法會導致局部最佳值，而非 絕對最佳值。 In computer music, music synthesis reduce the memory cost efficiently. Frequency-based analysis is a popular and easy way in computer music. The sepectra of most instruments aretime- varying. Synthesis matching procedure begins with a time-varying spectral analysis of the original music tone. Short-time Fourier analysis is needed to get the overall time-varyingspectra of an instrument tone. Wavetable synthesis model is a popular synthesis method because of efficiency and performance. In finding bases ( Wavetable ), We use four methods ( clustering, PCA, Group-additive and centroid-mapping ) to explore the feature of the time-varying spectra. Over 50% memory is reduced for each instrument tone and the synthetic tone sound indistinguishablefrom the original one. For multiple pitched tones of an instrument,We try to find a general set of wavetables to synthesize them. In the thesis, we compare these four methods with their performance and complexity. LBG algorithm is the easiest way tofind the bases and achieve a good performance but it results in local minimal problem.