- 學位論文
以非重疊離子佈植記憶元件類神經突觸模型模擬語音辨識系統之研究
Study of Simulating the Speech Recognition System by Non-overlapped implantation (NOI) MOSFET synapse Model
摘要
本論文以非重疊離子佈植(Non-overlapped implement, NOI)電晶體模型作為突觸單元，模擬一NOI 矩陣為960×16的類神經架構，驗證16組皆為兩個字的中文語音指令辨識系統。每組中文語音指令分別以20個樣本進行訓練，並且另外收集10個樣本進行辨識測試。語音樣本以梅爾頻率倒頻譜擷取語音特徵值，並藉由音框及帶通濾波器在時域和頻域對訊號作分析。 為了使元件模型能夠符合類神經演算法，故將所得的類比輸入轉為數位訊號進行訓練。系統經訓練1000圈雖然無法收斂，但仍可獲得69%的整體系統辨識率，成功利用模擬驗證了此系統在未來實現於硬體的可行性，未來預計將此語音辨識系統應用於智能家庭。 最後，本論文還構思出一擴充機制，在不影響原始系統的辨識率的狀況下，增加新的指令成為系統觸發指令。使用者可以利用此擴充機制依據各自的喜好而有不同的設定，如此便可賦予此語音辨識系統可擴充的彈性及客制化的潛力。
並列摘要
In order to verify the feasibility of the speech recognition system using non-overlapped implantation (NOI) MOSFET synapse model, we expanded the NOI array from 4×3 to 960×16 and sampled the voice signal consisting of 16 two-word commands in mandarin by Mel-Frequency cepstrum (MFCC). In this process, we analyzed the input signals not only in the Time Domain but also in the Frequency Domain; besides that, we converted the sampled analog signal to digital to simplify the neural network algorithms. However, after we trained 20 samples and tested each command by simulating with the Non-overlapped implantation (NOI) MOSFET synapse model, the system did not converge. Nevertheless, we still can achieve a 69% recognition rate. The result shows that the speech recognition realized system by the Non-overlapped implantation (NOI) MOSFET synapse model has a great potential and can be used in smart home. In addition, we also proposed an expanded feature for this neural network recognition system. This feature would make it more flexible by adding a new command based on the user’s preference.