本研究提出一套基於條件式自編碼器與 DiffWave 聲碼器之連續混合音色轉換系統，目標是以鋼琴音訊作為來源輸入，將其音色轉換為介於小提琴與二胡之間的可控制混合音色。系統首先將 WAV 音訊轉換為 CQT 幅度頻譜，並於條件式自編碼器中使用 log-CQT 幅度頻譜進行頻譜域音色轉換。不同樂器音色透過 style embedding 表示，並在小提琴與二胡之 embedding 空間中進行線性內插，以產生 30% Erhu + 70% Violin、50% Erhu + 50% Violin 與 70% Erhu + 30% Violin 等不同比例之混合音色，再由 DiffWave 聲碼器還原為 WAV 音訊。 實驗部分採用 FAD、ViSQOL、MRSTFT、CQT Loss、高頻音色特徵差異率、HNR 與 Estimated Noise Percentage 等指標進行評估。結果顯示，生成音色能依比例設定朝小提琴或二胡端點產生方向性變化，表示本研究方法具備初步的混合音色控制能力。然而，生成結果在高頻擦弦細節、二胡音色自然度與諧波穩定性上仍有改善空間。整體而言，本研究初步驗證跨文化拉弦樂器連續混合音色轉換之可行性，並可作為未來多音色融合與 AI 輔助聲音設計之研究基礎。;This study proposes a continuous hybrid timbre transfer system based on a conditional autoencoder and a DiffWave vocoder. The goal is to use piano audio as the source input and convert its timbre into a controllable hybrid timbre between violin and erhu. The system first converts WAV audio into CQT magnitude spectrograms， and then performs spectral-domain timbre transfer using log-CQT magnitude representations within the conditional autoencoder. Different instrumental timbres are represented through style embeddings， and linear interpolation is conducted in the embedding space between violin and erhu to generate hybrid timbres with different mixing ratios， including 30% erhu + 70% violin， 50% erhu + 50% violin， and 70% erhu + 30% violin. The converted representations are then reconstructed into WAV audio using the DiffWave vocoder. In the experiments， several evaluation metrics are adopted， including FAD， ViSQOL， MRSTFT， CQT Loss， high-frequency timbral feature difference rate， HNR， and Estimated Noise Percentage. The results show that the generated timbres exhibit directional changes toward either the violin or erhu endpoint according to the specified mixing ratios， indicating that the proposed method has an initial capability for hybrid timbre control. However， there is still room for improvement in terms of high-frequency bowed-string details， the naturalness of the erhu timbre， and harmonic stability. Overall， this study preliminarily verifies the feasibility of continuous hybrid timbre transfer for cross-cultural bowed string instruments， and may serve as a foundation for future research on multi-timbre fusion and AI-assisted sound design.
