隨著高齡化社會與長照需求增加，構音障礙患者之溝通困難受到重視。患者常因神經肌肉控制受損，產生語音模糊、語速不穩、音節拖長與韻律異常，使照護者難以及時理解需求，影響照護效率與安全。因此，提升構音障礙語音可理解度，對改善長照品質具有意義。本研究提出基於 Modified SpeechT5 與 Modified HiFi-GAN 之語音清晰化系統。Modified SpeechT5 加入 pe-bias 與 Feature Refinement Layer，以改善時間對齊偏移與頻譜模糊；Modified HiFi-GAN 導入 Snake Activation，以提升語音週期性與諧波結構重建能力。系統整合 Telegram Bot，使用者可透過手機傳送語音，由本地端伺服器完成前處理、推論、清晰化與結果回傳。為呼應永續綠能與節能減碳需求，本文建立推論階段能源消耗與碳排放估算模型，將 AI 模型推論時間、設備功耗與電力排碳係數納入分析，並比較不同模型維度與參數設定下之清晰化性能與推論時間，以達成兼顧患者溝通福利、長照通訊即時性與節能減碳需求之目標。;With the growth of aging populations and long-term care needs， communication difficulties among patients with dysarthria have become an important caregiving issue. Impaired neuromuscular control often causes blurred speech， unstable speaking rate， prolonged syllables， and abnormal prosody， making it difficult for caregivers to understand patients’ needs promptly and affecting care efficiency and safety. Therefore， improving dysarthric speech intelligibility is meaningful for long-term care quality. This study proposes a speech clarification system based on Modified SpeechT5 and Modified HiFi-GAN. Modified SpeechT5 integrates pe-bias and a Feature Refinement Layer to reduce temporal alignment deviation and spectral blurring， while Modified HiFi-GAN adopts Snake Activation to enhance speech periodicity and harmonic reconstruction. The system integrates a Telegram Bot， allowing users to send speech through mobile phones while a local server performs preprocessing， inference， clarification， and result delivery. To address energy-saving and carbon reduction requirements， this study establishes an inference-stage energy consumption and carbon emission estimation model using AI inference time， device power consumption， and electricity carbon intensity， balancing communication support， timeliness， and sustainable deployment in practical daily long-term care environments with limited computational resources.
