Diversification of Malay language database based on common voice features

Lee, Jing Wen (2020) Diversification of Malay language database based on common voice features. Project Report. Universiti Teknikal Malaysia Melaka, Melaka, Malaysia. (Submitted)

Text (Full Text) Diversification of Malay language database based on common voice features.pdf - Submitted Version Download (2MB)

Abstract

This project is about development of Malay language Databases with different voice profiles. The development is done by voice conversion, which will convert one source voice into different voice profile, enrich the Malay language databases. The objectives of this project are to develop voice conversion system to enrich the Malay language database with different voice profiles, to perform voice conversion using audio signal manipulation method and deep learning method with Deep Bidirectional Long Short-Term Memory based Recurrent Neural Network (DBLSTM) model and to analysis the converted voice audio based on human perception and waveform analysis. Two method of voice conversion is used which are phase vocoder and deep learning method. For phase vocoder method, the source audio is manipulated using time stretching and pitch shifting technique to produce different voice profile. Then the deep learning method will use CMU ARCTIC databases as the target output voice to train and develop a DBLSTM model. The trained model will be used to convert the source audio into the target voice that used in model training. Analysis will be made for the resultant output audio to determine the effectiveness and accuracy of the voice conversion system. The output result of phase vocoder can be observed that the linguistic information of the word can be preserved, but the sound quality does not meet the requirement of a normal human voice. On the other hand, the trained DBLSTM model can reproduce human voice in its converted speech, but the linguistic information of the words is only barely preserved. In conclusion, the voice conversion system developed can produce reasonable results and hopefully the system can be enhanced to perform better voice conversion.

Actions (login required)

View Item