- 學位論文
改善樣本分類與角度測量以增進個人化 HRTF 之合成
Improving the Classification of Samples and Measurement of Angle to Enhance the Synthesis of Personalized HRTF
摘要
近年來3D立體音效技術廣泛運用在各種應用，要產生逼真的3D立體音效，就必須探討音源定位的問題。由於每個人的外觀差異，例如軀幹、肩膀、頭型、耳廓大小的不同，造就出個人獨特的HRTF(Head Related Transfer Function)。獲得個人化HRTF的方法有很多種，但本文希望找出簡單且快速的方法來找出適合受測者的個人化HRTF，利用聆聽的方式找到個人化HRTF。 本文提出的演算法是先對CIPIC HRTF 的樣本進行分類，利用訊號特性細分資料庫，找出各角度的特徵代表頻率，再利用這些特徵代表頻率找出吻合的樣本，接著進行試聽測試。而本文的演算法，有別於過去研究需要試聽許多的樣本才能找到個人化的HRTF，可以降低受測者試聽的次數。 在做個人化HRTF驗證時，本文是使用感應器量測角度，可避免人為判讀的誤差以提高測量角度的精確性。 本文演算法與先前演算法比較後，發現在時間上可以縮短約為31%、角度誤差降低約32%與標準差減少約38%。因此，改善樣本資料的分類與利用感應器來測量角度可以有顯著的效果。
並列摘要
In recent years, the 3D audio effect is widely used in many applications. To produce 3D audio effect as real, the localization of sound source must be discussed. Because person's exterior is different to others, for example the torso, the shoulder, the head, and the auricle. These differences accomplish each person's unique HRTF (Head Related Transfer Function). Although there are many ways to personalize the HRTF, in this paper it is expected to find a easy and fast way to obtain the suitable personalized HRTF. That is a way to listen and synthesize the personalized HRTF. In this paper, an algorithm is proposed to synthesize the personalized HRTF by improving the classification of samples and the measurement of location angle. At first, classify the HRTF of samples from CIPIC, use signal characteristics of classified database, and find the representation frequency of each of angle. Then, the characteristic frequency is used to filter samples. The filtered samples are synthesized a HRTF that is used to evaluate by listening tests. The algorithm of this paper is different from previous researches that is needed to listen to many samples to find personalized HRTF. Therefore, the times of the listening test can be reduced. To verify the personalized HRTF, the angle of localized sound source is measured by sensor in this paper. The error due to human interpretation can be avoid to improve the accuracy of the measured angle. Compared with previous research, the time is shortened by approximately 31%, the angle of error is reduced by about 32%, and the standard deviation is also decreased by approximately 38%. Therefore, improving the classification of samples and measurement of angle can have a significant effect.