|
|
Text-to-Speech Personalization
Luner, Michal ; Černocký, Jan (oponent) ; Brukner, Jan (vedoucí práce)
This thesis aims to develop a model that can convert input text written in Czech into speech that closely resembles a target speaker. This work is based on the VITS text-to-speech neural network model. The workflow is as follows: a Czech dataset is acquired, the neural network is trained, the trained model is then used to generate audio samples, which are evaluated using several objective metrics. A personalized dataset is developed and used to fine-tune the model, and the evaluation process is repeated. As a result, two fine-tuned models were developed. The male model achieved a~MOS of 4.12, and the female model achieved a~score of 3.02. The scores prove that a base model fine-tuned using a personalized dataset can achieve results close to the original audio. The contribution of this thesis is, apart from the personalized models, the pipeline for audio evaluation and dataset development, which can be easily adjusted for tasks on different data. In addition, a detailed analysis of best practices applied during the development of new datasets is provided.
|
host ::
RSS.