New discrete diffusion model enhances speech inpainting and editing

Researchers have developed SIEDD, a novel discrete diffusion model designed for speech inpainting and editing. This framework, named HiCoDD, operates on hierarchical codec tokens, iteratively refining masked segments while maintaining speaker identity, prosody, and recording conditions. SIEDD demonstrates superior performance on the RealEdit benchmark for speech editing and outperforms autoregressive baselines in speech inpainting tasks, showcasing the benefits of explicitly modeling codec hierarchies for context-preserving audio reconstruction. AI

IMPACT This model could improve audio editing tools and content creation by enabling more precise reconstruction and modification of speech segments.

RANK_REASON The cluster contains a research paper detailing a new model and its performance on benchmarks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• HiCoDD
• Hugging Face
• RealEdit
• SIEDD

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →