Gefilte TTS
Personal Hebrew TTS voice from a non-native speaker's read sentences
About This Project
Gefilte-TTS is a text-to-speech project designed to synthesize a Hebrew voice modeled after a non-native speaker. The system automates the preparation of training data by processing Hebrew sentences from the Tatoeba dataset, applying nikud (vowel marks) via cloud-based computation, and correcting orthographic inconsistencies such as silent vav handling. The processed data is then curated and published to a Hugging Face dataset for model training.
The project utilizes a modular Python pipeline to manage the workflow from raw text acquisition to dataset preparation. By automating the normalization of Hebrew script, it streamlines the creation of high-quality speech synthesis datasets. This project is useful for researchers and developers interested in voice cloning techniques, particularly those working with non-native speech patterns or underrepresented linguistic datasets.