In this , we'll discuss the implementation of the Korean prosody generator using the back-propagation artificial neural networks. To make artificial neural networks learn the prosody residing in natural speech, we have to know what affects the pr...
In this , we'll discuss the implementation of the Korean prosody generator using the back-propagation artificial neural networks. To make artificial neural networks learn the prosody residing in natural speech, we have to know what affects the prosody in speech. On the basis of this knowledge, we designed the Korean prosody generator using back-propagation artificial neural networks that has shown the best performance for the Korean prosody generation. Back-propagation artificial neural networks for pitch period and energy contour were designed, trained and tested.
The input modules of each artificial neural networks consist of 11 phoneme strings that include synthetic symbols such as period, comma, and blank. Usually they are used as markers for the prosodic boundaries in a prosodic phrase.
To train and test these artificial neural networks, we made a corpus that consists of meaningful sentences or phrases which were generated by a group of phonetically balanced words.
Speech data read by a male speaker were recorded and collected as speech data. We analyzed these data to extract prosodic information of each phoneme, and made target and test patterns for artificial neural networks.
We trained artificial neural networks with target patterns and tested them with test patterns and could find out that artificial neural networks can learn the prosody in sentences well.
The estimation rates of pitch artificial neural networks were ranged from 91% to 92% and those of energy artificial neural networks were from 89% to 90%.
In this , we confirm that artificial neural networks can generate the prosodic information of sentences. In order to increase the estimation rates of artificial neural networks, we have to build up a large scale text corpus which include all the prosodic variations, and need more speech data. However, we should avoid the problem of over-training when the data for training was not sufficient.
，韩语毕业论文，韩语论文题目 |
