APLIKASI PENCIRIAN DENGAN TRANSFORMASI WAVELET UNTUK PENGENALAN PENGUCAP TEKS TAK BEBAS MENGGUNAKAN JARINGAN SARAF TIRUAN
| Main Authors: | Nugroho, Theodorus Yudho Dwi, Hidayatno, Achmad , Isnanto, R.Rizal |
|---|---|
| Format: | Thesis NonPeerReviewed application/pdf |
| Terbitan: |
, 2011
|
| Subjects: | |
| Online Access: |
http://eprints.undip.ac.id/32074/1/Theodorus_Yudho_Dwi_N.pdf http://eprints.undip.ac.id/32074/ |
Daftar Isi:
- The wavelet transform is a feature extraction method that can separate a signal into high and low frequency signals to obtain information that become the characteristics of speaker. Therefore, research is required to design a speakers recognition system and to analyze the type of motherwavelet and decomposition level which give a high percentage of recognition in the system. Type of mother wavelet and decomposition level are very important to extract the characteristics and information of a voice signal. In the previous research done by Leo Ignatius, wavelet transform is used for Indonesian vowel recognition application, whereas in this final project wavelet transform is used for speakers recognition application. The types of mother wavelet used are the Haar, Daubechies4, and Symlets2 wavelets with multilevel decomposition of 2, 4, and 8 which use backpropagation neural network as learning and recognition algorithm of the system. This research was done with the following steps. The first step is data acquisition process which records 10 respondents‟ voices, design the speakers recognition system using discrete wavelet transform to take the characteristics of voice signals, and finally implementation into an artificial neural network which is designed to be recognized as the speaker. Based on training data test results, the composition of mother wavelet and decomposition level which give the best recognition percentage is using wavelet Daubechies4 (db4) with two multilevel decomposition which is equal to 97%. While in the testing of test dependent text data, the best recognition percentage is wavelet Daubechies4 with two multilevel decomposition which is equal to 83.33%. Keywords: mother wavelet, decomposition level, backpropagation, feature extraction