小提琴一直以來都是許多人學習與演奏的樂器，有許多膾炙人口的 曲子與優秀的小提琴音樂家。在眾多曲子中，小提琴與其他樂器的合奏 曲子佔多數，因此需要其他樂器的演奏者一起合奏才能完整呈現曲子的 風貌。然而，由於時間或成本的因素，尋找長期合作的合奏者 (伴奏者) 並不是那麼容易，網路上的公開資源又多為混合音訊，合奏的效果不佳。 因此本研究針對最常見的小提琴與鋼琴的合奏方式來開發一套系統，此 系統可將混合音源中的小提琴與鋼琴音源分離，並使用分離音源追蹤現 場小提琴演奏，輸出鋼琴伴奏。 本研究旨在開發一套使用音源分離結果實現小提琴演奏追蹤的即時 音樂追蹤系統，我們設計了音源分離模組與音樂追蹤模組，在音源分離 模組方面，我們自行蒐集並建立一套新的公開整合資料集，用於訓練 Band-Split RNN 模型，並改進了模型的頻帶切割方法。在模型的評估上，我們使用訊號失真比來計算模型的分離效果，結果顯示模型在資料缺乏 與資料充足的情況下皆優於現有的基線模型，並證明頻帶切割方法的有 效性。在音樂追蹤模組方面，我們改進了線上動態時間規整演算法與貪 心向後對齊方法，重現了即時音樂追蹤模組的設計，並改良部分元件。 在實際的測試中，即時音樂追蹤系統展現了低延遲與精準追蹤的表現， 並在不同特徵的追蹤表現上保持了與離線追蹤相同穩定的追蹤效果。;The violin has long been a popular instrument for learning and perfor mance, with many well-known pieces and distinguished violinists. Among these pieces, ensemble compositions involving the violin and other instruments are predominant, requiring collaboration with other instrumentalists to fully present the musical piece. However, due to time or cost constraints, finding long-term ensemble partners (accompanists) can be challenging due to time or cost con straints. Online public resources often provide mixed audio, which does not yield good ensemble effects. Therefore, this research focuses on developing a system for the common violin and piano ensemble. This system can separate the violin and piano sources from a mixed audio source, track the violin’s per formance using the separated audio, and output the piano accompaniment. The goal of this research is to develop a real-time music tracking system that utilizes source separation results to track violin performances. We designed a source separation module and a music tracking module. For the source sep aration module, we collected and established a new open integrated dataset to train the Band-Split RNN model, improving the model’s band-split method. We evaluated the model using the Signal-to-Distortion Ratio to measure the separa tion performance. The results show that the model outperforms existing baseline models in both data-limit and data-rich cases, demonstrating the effectiveness of the band-split method. For the music tracking module, we improved the On line Dynamic Time Warping algorithm and the Greedy Backward Alignment method, reimplementing the design of the real-time music tracking module and enhancing some blocks. In practical tests, the real-time music tracking system exhibited low latency and accurate tracking performance, maintaining stable tracking results comparable to offline tracking across different feature tracking performances.
