Application of The Speaker Identification By Plastic Perception Neural Network

• Main Authors: Yu-Liang Jen, 任祐良
• Other Authors: I-Chang Jou
• Format: Others
• Language: zh-TW
• Published: 2003
• Online Access: http://ndltd.ncl.edu.tw/handle/23829078242056068654

碩士 === 國立高雄第一科技大學 === 電腦與通訊工程所 === 91 === The biometrics shall be proved to be a significant work of research in the future, and its methods are often being used by voiceprints, retinas, facial features, fingerprints, and fingered shapes. To use the image feature in identification, frequently has to use the equipment of image scanner to obtain the image data in need, and such equipments are quite expensive compared to the practical environment. Therefore, a daily application shall comparatively be impossible. However, the voiceprints need only a simple recording outfit that may have used the speaker identification necessary and that to be accepted by the public. The technique of the speaker identification itself involves a lot of related knowledge including notion of the digital signal processing, designs of the wave filter, basic knowledge of the phonetics, and the application know-how of the classifiers, etc. Although it seems not easy to get into an introductory study, the speaker identification is indeed a very interesting research item, either in application to recognize the status in network or in family entrance guard, in future, people needn’t to memorize the redundant and complex code number or to carry a string of keys instead of using the intrinsic voice as the basis to recognize the status. As to apply the neural network on the voice feature, there are many researchers pointing out and retaining in an affirmative attitude. Taking the plastic perception as a structure of the classifier, this research is to input data of the voice feature through the learning in plastic perception and calculating up a set of weighting value interactively responding to the speakers, with which the status recognition afterwards needs only inputted the voice feature and the weighting value in database and activated the recall memories in plastic perception, and then the speaker status will be recognized. The discrimination between good and bad voice identification systems is dependent of the rate and speed of its corrective identification, which is often influenced by the numbers of the speakers. In this thesis, the training and identification of the plastic perception can be computed by parallel network, and the speed is much faster than that of conventional neural network. From the experimental tesults, The optimum efficiency of error rate in inside test is 1.1% and the optimum efficiency of error rate in outside test is 1.1% on the speaker identification system (25 inside training speaker, 3 outside test speaker).