客製化小提琴指法生成系統

Abstract

指法的選擇對於小提琴演奏的流暢度和音樂性具有重要影響。為了協助小提琴初學者，現有方法主要依賴樂譜資訊，透過深度學習模型自動生成小提琴指法。然而，這些模型在生成指法時存在著限制，通常只能生成單一或特定種類的指法，無法滿足客製化需求。因此，我們提出了一種新的模型，結合小提琴演奏的音訊和樂譜資訊，使模型能夠根據不同演奏的錄音生成多樣化的指法。使用者可以通過上傳他們喜愛的演奏家錄音來學習該演奏家的指法。除了現有的資料集外，我們還在 YouTube 上收集了三十首含有其他樂器伴奏的小提琴演奏片段，以測試我們模型的表現。此外，我們嘗試使用自迴歸模型生成更加流暢的指法序列。該模型同時允許使用者手動調整特定音符的指法，並自動微調相鄰音符的指法，以提升整個序列的合理性。綜合以上所述，本研究通過引入小提琴演奏的音訊和使用自迴歸模型，旨在生成客製化的小提琴指法並提升模型的準確性。

The selection of fingering is crucial for achieving a fluid and expressive violin performance. To help novice violin learners, deep learning models have been developed to generate recommended fingerings from symbolic data. However, these models are not able to customize specific violin players’ fingering choices. To address this limitation, we propose a novel model that integrates both audio and symbolic data to generate fingerings based on a specific violin performance. Users can upload audio recordings of their favorite violinists to learn their fingering techniques. We also collected a new dataset from YouTube to evaluate the model’s performance under noisy conditions. Moreover, we introduce an autoregressive model that generates fingering based on context. Users can specify some fingerings, and the model will predict the fingerings for the remaining notes. Our proposed approach provides a promising solution to generate customized violin fingering and improve the overall model performance.