| 摘要: | 聲源定位是一個難以解決的工程問題，使用傳統的方法無法有效地找出聲源的位置。受益於硬體與訊號處理技術的進步，我們可以使用麥克風陣列來探測聲源的位置。隨著聲源定位的應用領域越來越多，發展了許多演算法，常見的有延遲相加波束成型(Delay-And-Sum Beamforming, DAS)、最小能量無失真響應(Minimum Power Distortionless Response Beamforming, MPDR)與多重訊號分類(Multiple Signal Classification, MUSIC)，近年來電腦計算能力增加，出現了利用去卷積的演算法如：Clean-SC。本研究比較DAS、MPDR、MUSIC與Clean-SC四種演算法運用在遠場聲源定位的性能。
研究中使用小型麥克風陣列，比較不同麥克風陣列幾何配置於200Hz至6000Hz的波束圖型，選用25個麥克風的螺旋陣列作為本研究的麥克風陣列幾何配置。單聲源模擬的結果顯示，DAS與MPDR所需的運算時間較少但精確度較差，MUSIC精確度高但花的時間大概是DAS與MPDR的兩倍，Clean-SC花的時間最多，大於1秒。雙聲源實驗考慮3種不同聲源間距：55、35與15公分，在聲源間距55公分2000Hz雙聲源模擬結果，DAS與Clean-SC無法定位出兩個聲源，在3000Hz與4000Hz，四種演算法皆能定位出兩個間距55公分的聲源。在聲源間距35與15公分，DAS皆無法分辨兩個聲源，故實驗選擇55公分作為雙聲源間距。
實驗在有迴響的房間進行，單喇叭實驗中，在2000Hz四種演算法皆能夠找到位於喇叭高音單體處的聲源，其中又以MUSIC與Clean-SC最為精確，但在3000Hz與4000Hz只有MUSIC能夠定位出一個聲源，其餘演算法受到桌面反射的影響，聲源位置實際還要低大約10公分。2000Hz雙喇叭實驗結果顯示DAS與MPDR無法分辨出兩聲源，MPDR在兩聲源間距為55公分下，因為主瓣重疊而無法分辨兩聲源，MUSIC能夠定位出兩個聲源的位置但有8公分的水平偏移，Clean-SC受到DAS演算法的結果影響而無法分辨兩聲源。在3000Hz只有MUSIC能夠約略判斷出兩個聲源，4000Hz所有演算法皆無法分辨出兩個聲源。在運算時間上，趨勢與模擬結果一致。
由模擬與實驗結果可以知道，小型麥克風陣列運用去卷積演算法Clean-SC於單聲源定位，能夠增加陣列的精確度，但是需要更多的運算時間，且容易受限於DAS的結果，在多聲源與有迴響的環境下，使用MUSIC較佳。研究中使用窄頻聲源定位演算法，未來希望能夠加入相關子空間法(Coherent Subspace Method)來偵測不同頻率的聲源。
;Sound source localization is a difficult problem in engineering. Traditional methods generally cannot locate sound sources effectively. Due to the development of technology in hardware and signal processing, the microphone array can be used for source identification.
Recently the application of sound source localization has grown rapidly, and many algorithms have been developed for it. Three common techniques are Delay-And-Sum Beamforming (DAS), Minimum Power Distortionless Response Beamforming (MPDR), and Multiple Signal Classification (MUSIC). As computing power grows exponentially, deconvolution algorithms, such as Clean-SC, have been developed. This study compares the performance of four algorithms, such as DAS, MPDR, MUSIC and, Clean-SC, in far-field sound source localization.
A small-sized microphone array was used in this study. By simulating the beam patterns of different array configurations between 200Hz and 6kHz, a well-performing spiral array configuration with 25 microphones was chosen. The simulation of single source showed that DAS and MPDR took less computational time but reduced the accuracy; MUSIC and Clean-SC were more accurate at the expense of computing. Much more computing time were required for Clean-SC than for MUSIC. Two source simulation considered three distance, 55cm, 35cm ,and 15cm, between two sources. In 55cm simulation, DAS and Clean-SC could not locate two sources at 2000Hz. Four algorithms were able to distinguish two sources at 3000Hz and 4000Hz. Because, DAS could not locate two sources 35cm and 15cm apart. 55cm was chosen for the distance between the two sources in the experiment.
The experiment was carried out in a reverberant room. At 2000Hz, the results revealed that four algorithms were able to find the position of a sound source locating at the tweeters of the speaker. Due to the overlapping of the main lobes, MPDR could not distinguish two sound sources. MUSIC was able to locate two sources but slightly deviated 8cm. Influencing by DAS, Clean-SC located the sound source incorrectly. At 3000Hz, only MUSIC was able to locate two sources but with lower accuracy. However, for 4000Hz, four algorithms could not locate two sources.
Finally, the results of simulation and experiment of a small array showed that Clean-SC, based on deconvolution, can achieve higher accuracy, but exchanged for computational time, compared to DAS, MPDR, and MUSIC in single source localization. MUSIC has better performance for multiple sound sources in reverberant environments. The narrow-band algorithm is used in this research. In the future, it is hoped to apply Coherent Subspace Method to detect sound sources with different frequencies. |
