In the thesis, we addressed the problem of Slovene speech synthesis based on relatively small data set. We described older approaches to speech synthesis like articular and formant synthesis, and more modern approaches like unit selection and speech synthesis with deep neural networks.
We created a dataset consisting 30 hours of speech from four speakers for use with
speech synthesis. We used ForwardTacotron architecture for generating mel-spectrograms and Hifi-GAN architecture for generating waveforms from these spectrograms. We created a basic model for male speech, which can be fine-tuned for new speakers. The best system we created achieved a good mean opinion score of listeners (4.07 on a scale
1-5) that simulates natural speech.
|
