粒子群最佳化支撐向量機結合改進適應度函數於語者辨識之研究

Study on Speaker Recognition by combining Particle Swarm Optimization of Support Vector Machine with Improved Fitness Function

古欣亞(Hsin-Ya Ku)

指導教授 ： 陳國在

國立臺灣大學/工學院/工程科學及海洋工程學研究所/碩士(2012年)

https://doi.org/10.6342/NTU.2012.10713

摘要

隨著科技的進步，以往以密碼確認身分的方式已經不再是一種高安全性且方便的保護，而生物辨識技術也隨著科技的進步慢慢茁壯，在所有的生物特徵中，語音可以說是人類身上最為自然的特徵之一，代表著如果研發出一套辨識率高且方便的系統，將可帶給人們更便利的生活。

在眾多的機器學習方法中，建立在統計學習理論上的支撐向量機是一種分類性能以及推廣能力都非常強的方法，數年來許多語者辨識的研究都是建立在支撐向量機上，只要參數設定良好的話，通常可以達到很好的辨識能力。

不過以往支撐向量機的最佳化過程中，都是以交叉驗證的方式作為進化算法的適應度函數，本文認為僅僅利用交叉驗證是一種不太可靠的方法，因此提出一種改進的適應度函數，配合梅爾倒頻譜係數將其應用到語者辨識系統中，以期望獲得更高的辨識率，實驗結果顯示，本文提出改進的適應度函數結合粒子群演算法比以往僅利用交叉驗證辨識率作為適應度函數的方式效能得到了提升。

關鍵字

支撐向量機 ； 粒子群演算法 ； 梅爾倒頻譜係數 ； 適應度函數 ； 語者辨識

並列摘要

Because of high development of scientific technology, identifying people by code is never highly secure and conveniently protective. Meanwhile, biometric recognition technique is also much developed to replace the past recognition technique. To human being, speech is of biometrics and has the most natural feature. It does mean to bring human more convenient life when a highly accurate and convenient system of recognition is developed.
Among all machine learning techniques, support vector machine by basing on statistical learning theory is a very powerful one for both classifying and predicting data. For several years, many studies on the speaker recognition were based on support vector machine. It is usually able to get good accuracy if the parameters involved were set well.In the optimization process for past support vector machines, using accuracy of the cross validation as fitness function is commonly adopted. However, this study think that only using the accuracy of the cross validation as fitness function is not reliable. Therefore, it is to propose an improved fitness function supposed to get high accuracy. Consequently compared to those past methods, the accuracy had been promoted by using the improved fitness function.

並列關鍵字

Support Vector Machine ； Particle Swarm Optimization ； Fitness Function ； Mel-scale Frequency Cepstral Coefficients ； Speaker Recognition

參考文獻

1. B. S. Atal, Automatic recognition of speakers from their voices. Proceedings of the IEEE, 1976. 64(4): p. 460-475.

2. R. J. Mammone, Zhang Xiaoyu and R. P. Ramachandran, Robust speaker recognition: a feature-based approach. Signal Processing Magazine, IEEE, 1996. 13(5): p. 58.

3. A. E. Rosenberg, Automatic speaker verification: A review. Proceedings of the IEEE, 1976. 64(4): p. 475-487.

4. H. Gish and M. Schmidt, Text-independent speaker identification. Signal Processing Magazine, IEEE, 1994. 11(4): p. 18-32.

5. J. P. Campbell, Jr., Speaker recognition: a tutorial. Proceedings of the IEEE, 1997. 85(9): p. 1437-1462.