邊緣或手持裝置逐漸普及和海量成長，造成雲計算（Cloud Computing）的負載過大，延遲增長、可靠性降低等等問題。但隨著 晶片技術、物聯網（Internet of Things）技術的進步和提倡邊緣運算 （Edge Computing）概念的提倡，越多的邊緣裝置已經可以處理、執 行簡單的機器學習任務或程式執行，將大多機器學習和其他工作從雲 端下放至邊緣裝置上執行。 本論文將用樹莓派（Raspberry Pi）結合邊緣運算理念實現一語者 驗證系統。集中一地錄音和結合 One-versus-Rest 分類策略訓練一卷積 類神經網路（Convolutional Neural Network）模型，將模型部署到邊 緣裝置：樹莓派上，樹莓派即可原地端完成語者驗證的工作，無須交 由雲端或伺服器端負責機器學習辨識預測的工作。再將預測結果紀錄 於 Microsoft Azure 雲端資料庫上，以利日後使用。;Due to growing popularity and rising quantity of handheld and edge devices, traditional Cloud Computing are facing overloading issues, such as increasing reaction time, low reliability, etc. Now, more and more edge devices can handle simple Machine Learning tasks and able to execute programs locally instead of the Cloud with the benefit of semiconductor chips technology evolution, Internet of Things improvement and the advocations of Edge Computing Theorem. This paper implements a Speaker Verification framework using Raspberry Pi in Edge Computing. Record audio and create a Convolutional Nerul Network model combining Ove-versus-Rest strategy at one site. Deploy the model to our edge device: Raspberry Pi. It can execute our classification task on site rather than letting the cloud or server do the work. We then log the predictions on our Microsoft Azure Cloud SQL database for further usages.
