HIGH-FIDELITY TEXT-TO-SPEECH MODELS WITH SYNTHETIC ANNOTATIONS

Summary from the TLDR AI newsletter

These text-to-speech models, trained by Stability AI, can be guided by precise natural language instructions. As there is no large dataset with proper textual descriptions of audio for generation, its creators synthetically annotated a large corpus of speech for training. This is another example of a broader trend of annotation, up-captioning, and training for generative modeling.