New AI Dataset Lets You Edit Speech Like Text—Add Laughs, Remove Coughs
This could make podcast and video editing as easy as editing a document.
Researchers introduced VoxCanvas-41M, a large-scale bilingual dataset of approximately 41 million speech editing pairs covering content editing—single-span, multi-span, and location-reference editing—and paralinguistic editing, including insertion and deletion of 39 classes of nonverbal events such as laughter, sighs, and coughs. Building on VoxCanvas-41M, they trained VoxCanvas-Edit, an instruction-based speech editing model, and systematically investigated how models learn different editing operations and how data scale, task interactions, and training strategies affect editing performance. The dataset, pipeline, and demos are publicly available.
- A new dataset with 41 million speech editing examples helps AI learn to edit audio by instruction.
- The model can add or remove sounds like laughter and coughs, making audio editing easier.
- The dataset and model are publicly available, which could lead to cheaper, simpler editing tools.
Why It Matters
This could make audio editing as easy as editing text, saving time and money for content creators.