New autoregressive model enables real-time speaker extraction

Researchers have developed a new autoregressive model for streaming target speaker extraction, addressing the limitations of current generative models that struggle with real-time applications. Their approach, the Chunk-wise Interleaved Splicing Paradigm, ensures stable and efficient streaming inference by using historical context to refine extracted speech segments and mitigate discontinuities. Experiments on Libri2Mix demonstrate that this method maintains stability and superior intelligibility, even surpassing offline baselines at low latencies, achieving a Real-Time-Factor of 0.248 on consumer GPUs. AI

IMPACT Enables real-time applications for speaker extraction, potentially improving voice assistants and transcription services.

RANK_REASON The cluster describes a new research paper detailing a novel model and methodology for a specific AI task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv
• Chunk-wise Interleaved Splicing Paradigm
• Hugging Face
• Libri2Mix
• Shuhai Peng

• paper
• model release

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