電子喉是無喉者常用的替代發聲裝置，但其語音音質粗糙、可理解度偏低，且伴隨持續的機械噪聲。電子喉語音轉換（Electrolaryngeal Speech Voice Conversion，ELVC）旨在將電子喉（EL）語音轉換為自然（NL）語音，以協助無喉者更自然地溝通。然而 EL 語料在現實中極為稀缺，使得從頭訓練深度神經網路並不可行，而現成的任意語者轉換模型亦因 EL 語音屬於異常輸入（out-of-distribution）而難以直接套用。本研究即以資料稀缺為核心問題，提出並比較四類方案。 其一，以範例式最近鄰居法為基礎提出改良版 LLE-ELVC，將鄰居搜尋與權重估計解耦，分別交由語意特徵與韻律特徵負責，並開放語意特徵抽取模型的替換空間，在不增加資料需求的前提下提升轉換品質。其二，利用改良架構執行反向轉換以合成內容對齊的語料，並探索以 HiFi-GAN 直接學習 EL 至 NL 波形映射的可行性。其三，引入流對齊（Flow Matching）框架，將 EL 與 NL 特徵分佈之間的轉換建模為跨域映射問題，並透過序列層級的上下文建模改善韻律連貫性。其四，設計具多條件控制的生成式模型 MelSynthFlow，先以大規模語料進行生成式預訓練，再針對 ELVC 任務微調，以緩解資料稀缺對模型能力的限制。 實驗以 TMHINT 台灣華語平行語料為核心資料集，並輔以 LibriSpeech 與 COSPRO 進行聲碼器與生成式預訓練。客觀評估顯示，改良版 LLE-ELVC 相較於 kNN-VC 與基線方法在可理解度與頻譜失真上均有改善；流對齊模型在頻譜映射與韻律建模上較端對端波形映射更具優勢；MelSynthFlow 則在語音自然度上表現最佳，惟仍存在發音錯誤與語速適應不足等問題。整體而言，本研究系統性地比較了上述方法在低資料量條件下的表現與限制，可作為後續 ELVC 研究之參考。;The electrolarynx is a common alternative phonation device for laryngectomees， but the speech it produces is rough in quality， low in intelligibility， and accompanied by persistent mechanical noise. Electrolaryngeal speech voice conversion (ELVC) aims to convert electrolaryngeal (EL) speech into natural (NL) speech， helping laryngectomees communicate more naturally. In practice， however， EL corpora are extremely scarce， which makes training deep neural networks from scratch infeasible， while off-the-shelf any-to-any voice conversion models cannot be applied directly because EL speech constitutes an out-of-distribution input. Taking data scarcity as its central problem， this thesis proposes and compares four families of approaches. First， building on the exemplar-based nearest-neighbor framework， an improved LLE-ELVC method is proposed that decouples neighbor search from weight estimation， assigning them to semantic and prosodic features respectively and opening up the choice of semantic feature extractor， thereby improving conversion quality without increasing data requirements. Second， the improved architecture is used to perform reverse conversion to synthesize content-aligned corpora， and the feasibility of learning a direct EL-to-NL waveform mapping with HiFi-GAN is explored. Third， a flow matching framework is introduced that casts the transformation between the EL and NL feature distributions as a cross-domain mapping problem， improving prosodic coherence through sequence-level contextual modeling. Fourth， a generative model with multiple conditioning controls， MelSynthFlow， is designed and first pre-trained on large-scale corpora before being fine-tuned for the ELVC task， in an attempt to alleviate the fundamental limitation that data scarcity imposes on model capacity. Experiments use the Taiwanese Mandarin TMHINT parallel corpus as the core dataset， supplemented by LibriSpeech and COSPRO for vocoder and generative pre-training. Objective evaluation shows that the improved LLE-ELVC yields gains in both intelligibility and spectral distortion over kNN-VC and the baseline; the flow matching model offers clearer advantages in spectral mapping and prosodic modeling than the end-to-end waveform mapping; and MelSynthFlow achieves the best naturalness， though it still exhibits pronunciation errors and limited adaptability to speaking rate. Overall， this thesis provides a systematic comparison of the performance and limitations of these methods under low-resource conditions， serving as a reference for future ELVC research.
