根據2026 年初衛生福利部統計資料，台灣登記在案之聽覺機能障礙者已超過十四萬人，而台灣手語（Taiwan Sign Language， TSL）為其日常溝通的重要媒介。然而，現有網路手語資源多以單詞查詢或固定句型影片為主，當使用者欲表達完整中文句子時，常需逐詞查找並自行組合對應手語影片，導致資訊零散且缺乏連續性。為提升手語數位內容之可用性，本研究建置一套由中文文字輸入至虛擬人手語動畫輸出的模組化系統，將中文至手語詞彙序列轉換、手語動作資料檢索、動作參數串接與虛擬人渲染整合於同一流程中。 在系統設計上，本研究將手語動作資料建置與使用者動畫生成分為兩類流程。手語動作資料建置流程使用EHM-Tracker 將既有手語影片轉換為包含人體姿態、手部動作與臉部資訊之參數化追蹤資料，並建立可由gloss 查詢的詞彙索引；使用者動畫生成流程則以前端中文至自然手語詞彙序列轉換模組（Sign Language Translation， SLT）產生gloss sequence，依序查詢並檢查對應之詞彙追蹤資料。系統接著透過motion envelope 分析擷取各詞彙之有效動作區間，視設定對詞彙參數進行時間平滑，並於相鄰詞彙片段之間加入過渡幀，以降低多詞彙串接時的姿態跳動與動作不連續問題。最後，系統將串接後的連續動作參數輸入GUAVA avatar rendering 模組，產生具人物外觀一致性之虛擬人手語動畫。 實驗部分，詞彙層級平滑實驗顯示，10 個樣本之SMPL-X body pose 與MANO hand pose normalized jerk cost 平均分別降低99.90% 與99.87%，表示高頻動作變化明顯減少；然而，hand-only PCK@0.10 平均由0.9476降至0.8868，顯示平滑度與手部逐幀對齊精度之間仍需取得平衡。本研究亦比較EHM-Tracker 於100-step 與1000-step 最佳化設定下之追蹤結果。結果顯示，1000-step 在雙手清楚可見樣本中通常能改善reprojection error 與PCK，但在單手低可見度或手部離開畫面情境下，較長最佳化可能造成誤追蹤風險。因此，本研究採用較保守之100-step 設定建立手語動作資料庫。此外，本研究透過使用者問卷評估25 支系統生成影片，共回收53 份有效問卷；四項指標平均分數介於3.80 至4.01 之間，其中準確度與手勢清晰度表現較佳。問卷結果顯示，本系統已具備中文至虛擬 人手語動畫生成之可行性，但在手部細節、臉部表情與句子層級自然度上仍有改善空間。整體而言，本研究證明結合SLT、既有手語影片資料、EHM-Tracker 與GUAVA 之模組化流程，可作為後續聽障友善溝通、手語學習與手語數位內容生成之基礎。;According to statistics released by Taiwan＇s Ministry of Health and Welfare in early 2026， the number of registered individuals with hearing impairment in Taiwan has exceeded 140，000， and Taiwan Sign Language (TSL) serves as an important medium for their daily communication. However， most existing online sign language resources are organized as isolated word lookup systems or fixed sentence videos. When users wish to express complete Chinese sentences， they often need to search for corresponding sign videos word by word and combine them manually， resulting in fragmented and discontinuous information. To improve the usability of digital sign language content， this study develops a modular system that converts Chinese text input into virtual human sign language animation by integrating Chinese-to-gloss conversion， sign motion data retrieval， motion parameter concatenation， and avatar rendering. The system design separates sign motion data construction from user-requested animation generation. During sign motion data construction， EHM-Tracker converts existing sign language videos into parameterized tracking data containing body pose， hand motion， and facial information， and a gloss-based index is established for motion retrieval. During animation generation， a front-end Chinese-to-natural-sign-language translation module， namely the Sign Language Translation (SLT) module， generates a gloss sequence from Chinese input， after which the system retrieves and validates the tracked motion data corresponding to each gloss. Motion envelope analysis is then applied to extract the effective motion interval of each sign word. Optional temporal smoothing is applied to word-level motion parameters， and transition frames are inserted between adjacent sign segments to reduce pose discontinuities and motion jumps during multi-word concatenation. Finally， the concatenated motion parameters are passed into the GUAVA avatar rendering module to generate virtual human sign language animation with consistent character appearance. In the experiments， word-level smoothing was evaluated on 10 samples. The normalized jerk cost of SMPL-X body pose and MANO hand pose decreased by an average of 99.90% and 99.87%， respectively， indicating a substantial reduction in high-frequency motion variation. However， the average hand-only PCK@0.10 decreased from 0.9476 to 0.8868， revealing a trade-off between motion smoothness and frame-level hand alignment. This study also compares EHM-Tracker results under 100-step and 1000-step optimization settings. The results show that 1000-step optimization generally improves reprojection error and PCK when both hands are clearly visible. However， in singlehand samples with low visibility or hands moving out of frame， longer optimization may increase the risk of mistracking. Therefore， this study adopts the more conservative 100-step setting for constructing the sign motion database. In addition， a user questionnaire was conducted to evaluate 25 system-generated videos， with 53 valid responses collected. The average scores of the four evaluation metrics ranged from 3.80 to 4.01， with accuracy and gesture clarity receiving relatively higher scores. The questionnaire results indicate that the proposed system is feasible for generating virtual human sign language animation from Chinese text， while hand details， facial expressions， and sentence-level naturalness remain areas for future improvement. Overall， this study demonstrates that a modular pipeline combining SLT， existing sign language video resources， EHM-Tracker， and GUAVA can serve as a foundation for future applications in accessible communication， sign language learning， and digital sign language content generation.
