Synthesizer for Serbian and English language based on neural networks has been developed


 

There are two basic groups of speech synthesis methods based on the text - methods based on the direct concatenation of speech segments and parametric method. Parametric methods, including those based on neural networks, have recently become even more popular, primarily thanks to their flexibility. Namely, they provide the possibility of changing the voice characteristics, that is, the identity of the speaker and the manner of speech, and as such, have led to the development of new, attractive applications. The AlfaNum team, using its own speech and language resources, software modules as well as certain open-source tools for neural network management, has succeeded to develop a high-quality parametric synthesis of speech in Serbian and English. The synthesis is highly comprehensible and exceptionally natural sounding and you can hear for yourself by listening to the following samples:

  • A speaker reads text:

    1.

    2.

    3.

    4.

    5.

    6.

  • A synthesized voice reads text:

    1.

    2.

    3.

    4.

    5.

    6.

Published 10.03.2017.