以CNN為基礎之語音辨識系統及應用於兩輪平衡車的控制

Abstract

本論文實現語音辨識及使用語音控制於兩輪平衡車，語音辨識系統使用基於TensorFlow之上執行的Keras完成，語音訊號利用梅爾頻率倒譜係數(Mel-Frequency Cepstral Coefficients, MFCCs)提取特徵值，並使用卷積類神經網路(Convolutional Neural Network, CNN)進行學習及建立模型。 兩輪平衡車使用Arm Cortex M0之微控制器實現，整體架構包含馬達、驅動電路、改變重心的機械結構及各類感測器。本論文採用比例、積分及微分控制器(Proportional-Integral-Derivative, PID)進行對兩輪平衡的控制，並以機械結構使重心改變達成兩輪平衡車前進或後退的功能。 本論文針對語音辨識系統架構修改進行實驗，挑選出正確率最高的架構應用於兩輪平衡車之控制中，最後實驗證實此論文的可行性。

This paper realizes a speech recognition system and voice control for a two-wheel balance vehicle. Based on TensorFlow, the speech recognition system is implemented by using Keras. The feature values of voice signals are extracted by using mel-frequency cepstral coefficients (1s), and then convolutional neural network (CNN) is used to learn and build a speech recognition model. The control of the two-wheel balance vehicle is accomplished by an Arm Cortex M0 microcontroller. The overall architecture consists of a motor, a drive circuit, various types of sensors and a mechanical structure that changes the center of gravity of the two-wheel balance vehicle. This paper uses PID controller to control the balance of vehicles. And the mechanical structure makes the center of gravity change in order to make the two-wheeled balance vehicle move forward or backward. This paper conducts some experiments on the modification of the speech recognition architecture, and selects the best recognition architecture to implement the voice control of the two-wheel balance vehicle. Finally, the experiment confirms the feasibility of the proposed method.