音樂學習智能虛擬教練：以小號音色為例

本研究開發小號的虛擬教練，將平常的基本練習作為教材，利用機器學習的技術，使用Unity 3D整合成樂譜的方式呈現在使用者的眼前，使小號的個人練習更有效率。
本研究著重在此小號虛擬教練的音色辨識上，透過時頻譜的分析，整理出相對正確的小號音色。用TensorFlow分析專家的音色、收集使用者的資料，並判別其相似度。透過數據化系統，使樂器學習更有效率、音色之準確度判斷更為準確。
機器學習可說是當今的熱潮，許多行業都結合此技術進行開發及應用，但目前在音色的辨識上應用較少。以往的音色辨識多是透過聲音判斷樂器的種類，或是透過鳥鳴聲判斷鳥隻的類別，利用時頻譜分析辨識音色的優劣的系統較少，本研究蒐集了專家的音色，透過傅立葉轉換將音頻轉換成圖像，使抽象的音色概念可視化，在樂器學習的部分能夠更加具體。
實驗結果顯示，泛音的多寡確實能夠影響音色，泛音越多則聽起來越飽滿、越接近專家的音色，使用者與專家的音色判別準確度均超過70%，在音色辨識上，時頻譜圖與圖像分析的結果能夠作為音色是否符合標準的的判斷要素之一。

During the process of practicing a musical instrument, basic exercise is considered as a convention of every single time of practicing, regardless of a beginner or a profession player. As one must has the steady foundation to proceed further. It is a common issue for beginners, even expert players who have been playing for a period of time, to ignore the importance and priority of basic exercise, and the correctness of the sound they performed. This study utilized technics of Machine Learning to collect learners’ data before and after learning, along with the evolving status of the smart virtual trainer. Integrated through Unity 3D development platform, in order to make musical instrument learning more efficient through the digitized system. This study is spent on the timbre identification, through the spectrogram analysis to find the correct timbre.
Machine learning is viral these days, but there isn't a lot about the timbre identification. Most of it is about to recognize what the musical instrument is, or identify what kind the bird is through its tweet. It is rare to judge the timbre is good enough or not.
This study collected some experts' timbre, and made the sound to be seen through the Fourier transform. To make the abstract timbre concept more specific.
According to the experimental results, it proved that the quantity of overtone is related to the timbre. According to the students’ testing results, the similarity between they and the experts are over 70%. Besides, spectrogram and the result of image recognition will be two of factors to affect the resolution.