本論文以兩個麥克風組成麥克風對建構出一個聲源方位偵測系統，其運算的部分是由嵌入式系統(TMS320VC5509)來完成，運算的部分包括語音活動偵測(Voice Activity Detection, VAD)、時間延遲(Time Delay of Arrival, TDOA)估測與方位角度估測，其中語音活動偵測結合了對數能量分析(Log Energy)與頻譜亂度分析(Entropy)來增加判定的準確性與降低總計算量。時間延遲估測的部分使用廣義交互相關函數(Generalized Cross Correlation, GCC)，並以拋物線內插法來增加時間延遲估測的準確性。在方位角度估測是依據麥克風對與聲源所建構出的雙曲線來求得，在估測中以此雙曲線的漸進線來作為聲源的方位角度。 而本論文也在實驗的部分對於時間延遲估測的幾種方式像是平均振幅差函數(Average Magnitude Difference Function, AMDF)、最小平方法(Least Median of Square, LMS)、交互相關函數(Cross Correlation, CC)與廣義交互相關函數做出討論與比較。In this thesis, a sound source localization system is studied and implemented. For the hardware part, the microphone array composed of two microphones is used to input the voice signals. The operation of this system is used by an embedded system (TMS320VC5509A). For the software part, VAD (Voice Activity Detection) and TDOA (Time Delay of Arrival) estimation and direction the detection are executed in order. In the processing of VAD, we combine the log-energy and the spectral-entropy to distinguish the speech/non-speech frames. To estimate the TDOA, an approximation algorithm is used to compute a generalized cross correlation function. Then, use the parabolic interpolation based method to increase the accuracy of estimated TDOA values. The TDOA and speed of sound in air can be employed find the direction of a sound source.
