In this page, we list some of the specific research topics we are working on.

在這個頁面，我們列舉一些我們正在做的具體的研究主題並簡單做出介紹。

• Symbolic music structure analysis 符號音樂結構分析

Symbolic music is a highly structuralized type of data, which involves multiple hierarchy levels such as note, chord, motif, phrase, subsection, section, etc. Symbolic music structure analysis aims to automatically analyze symbolic music, such as identifying repetitions (e.g., verse->chorus->verse->chorus) and the development of musical tension/emotion, which reveal the high-level compositon ideas of the composer. It also helps tasks such as music summarization and music captioning.

符號音樂（例如樂譜、MIDI檔等等）是一種高度結構化的音樂，包含了多種結構等級例如音符、和弦、音樂動機、樂句、樂段等等。符號音樂結構分析專注於自動分析符號音樂中的這些結構，例如辨識音樂當中的反覆（如主副歌的反覆）以及音樂情感或張力的發展。透過這類的分析可以進一步理解作曲家的作曲想法。從實務的角度上，也可以協助其他題目，例如自動對音樂進行總結或說明（music summarization/captioning）、音樂推薦系統等等。

• Music transcription 音樂轉譜

Music transcription focuses on automatically creating the transcription of a piece of music performance. Basically, this task is similar to developing a wav-to-MIDI conversion software! This is a fundamental task that helps various applications, from music learning (automatically generating musical score from audio), music understanding (such as determining the key and scale of the music), to music captioning systems.

音樂轉譜的目標是自動從音樂當中轉寫出原始的樂譜。簡單來說，這個問題相當於建立一個將wav轉換為MIDI檔的軟體。做為一個音樂研究上最基本的題目之一，音樂轉譜的結果有許多的應用，從音樂學習（自動從音樂當中產生樂譜）、音樂理解（例如辨識音樂調性）到更完整的自動對音樂進行描述、介紹、說明的系統，可說是極其重要的研究。

• Multimodal music semantic analysis 多模態音樂語意分析

Music semantic analysis aims at analyzing the semantic of music, which involves the study of how different parts of music (e.g., melody, lyrics, chord, rhythm) create meaning and convey emotions, tensions, or concepts. We are particularly intersting in multimodal music semantic analysis, which may involve different types of input, such as music audio, musical score, and lyrics. The use of multimodal data allows a more holistic analysis of music semantic.

音樂語意分析的目標是分析音樂的語意，即音樂的各個部分（如主旋律、歌詞、和弦、節奏等等）如何傳達出音樂的情感、張力、或音樂的概念。我們特別對於多模態的音樂語意分析感興趣。這類研究包含使用多種形式的輸入資訊，例如音樂訊號、樂譜、歌詞等等的資料，以對音樂作品做出更完整的分析。

• Automatic lyrics alignment and transcription 自動歌詞對位與辨識

Lyrics transcription recognizes the lyrics that the singer sings, while lyrics alignment focuses on providing the alignment between lyrics and sung audio, i.e., the timings where each character/word starts and ends. This is particularly important in automatic karaoke content generation! If you are interesting in karaoke, you may want to work on this topic. Of course, lyrics alignment and transcription also helps music semantic analysis by automatically transcribing the lyrics from music.

歌詞辨識的目標是自動辨識出歌手唱的歌詞（雖然這個解釋沒什麼意思XD），而歌詞對位的目標則是給定已知的歌詞與音檔，自動將歌詞與音檔對齊，即自動找出歌手演唱每個字/詞的起始時間與結束時間。這兩項研究對於自動產生卡拉OK的字幕可說是極其重要的研究！如果你對卡拉OK有興趣，你可能會想專注於這個題目。當然，歌詞辨識與歌詞對位也會對前面提到的音樂語意分析有幫助，因為它們可以自動從音樂當中抽取出歌詞的相關資訊。

• (Misc) Music analysis of various types of music 對於不同種類音樂的分析

This type of research involves specific types of music, such as Taiwanese Opera, classical music, Hakka music, Japanese music, etc. If you are a fan of some specific types of music and are interested in what really makes these types of music special (from other types of music), feel free to discuss with me. We can use computer science technology to analyze these music genres quantitatively in order to have a deeper understanding on such a music and culture.

這部分包括一些比較特定種類的音樂，例如歌仔戲、古典音樂、客家音樂、日本音樂等等。任何類型的音樂都在這個範圍之內。如果你有特別喜歡的音樂，想要知道這些音樂的特別之處是什麼，歡迎提出你的想法，我們可以用電腦科學的方式做出相關的定量分析，以對於這些音樂與文化有更深層的認識。