為了提升單一麥克風語者辨識系統的效能。本研究因此設計一個基於麥克風陣列的嵌入式語者辨識系統，系統分成四個模組：麥克風陣列聲音訊號擷取、波束成形、語者特徵擷取和語者辨識模組。聲音訊號模組經由微機電(MEMS)麥克風組成的環形麥克風陣列收集語者聲音資訊；波束成形模組藉由多通道聲音處理來增強語音訊號與除去周圍的雜訊；在語者特徵擷取模組，我們使用線性預測編碼倒頻譜(LPCC)來表示語者的聲音特徵模型；最後藉由機率神經網路(PNN)分類器來進行語者辨識。我們建置一個實驗的語者聲音資料庫，錄製十二人共120個相同語句的聲音資料，來驗證此一語者辨識系統，實驗過程藉由機率神經網路平滑參數與波束成形參數的訓練來最佳化辨識率。實驗結果顯示，基於麥克風陣列的語者辨識系統，相較於單一麥克風的語者辨識系統，可降低約百分之十的錯誤相等率。;The study is to design an embedded speaker identification system based on microphone array in order to improve the efficiency of single microphone identification systems. The system is composed of four modules including sound signal extraction from microphone array, beam forming, speaker features extraction and speaker identification module. Sound signal module is to collect speaker sound information by using loop microphone array composed of Micro Electro Mechanical System (MEMS) microphone; Beam forming is to enhance sound signal and remove background noise via multi-channel sound processing; Linear Predictive Cepstrum Coefficient (LPCC) is applied to represent a speaker sound characteristics module; The classifier of Probabilistic Neural Network (PNN) is applied to identify speaker. Besides, we built a database of experimental speaker sounds with one hundred and twenty same statements recorded by twelve people. This is to validate the speaker identification system. The recognition rate was optimized by PNN smoothing parameters and beam forming parameters during the training. The test results showed that our speaker identification system based on microphone array could reduce about 10% error rate compared to the single one.
