METHOD AND SYSTEM FOR TEXT-TO-SPEAK SYNTHESIS AND METHOD FOR TRAINING A TEXT-TO-SPEAK SYNTHESIS SYSTEM
Patent Information
- Application Number
- DE602021028970
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2020-08-28
- Filing Date
- 2021-08-27
- Publication Date
- 2025-04-09
- Estimated Expiration
- 2041-08-27
AI Technical Summary
Existing text-to-speech (TTS) synthesis systems struggle to generate speech that effectively conveys speech attributes such as emotion, intention, projection, pace, and accent, and they face challenges in training models using limited datasets while avoiding overfitting.
A method for training a text-to-speech synthesis system that involves obtaining two sub-datasets with varying levels of speech attribute prominence, sequentially training models using these datasets to gradually increase the speech attribute strength, and selecting the best prediction network for generating speech with desired attributes.
This approach enables the generation of speech that is perceived as natural, realistic, and human-like, while effectively training models with limited data and maintaining performance by avoiding overfitting.