Reproduce “Covo-Audio Technical Report” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wenfu Wang, Chenxing Li, Liqiang Zhang, Yiyang Zhao, Yuxiang Zou, Hanzhao Li, Mingyu Cui, Hao Zhang and 18 more
Tencent's Covo-Audio, a 7B-parameter end-to-end Large Audio Language Model (LALM), unifies continuous audio processing and generation to address limitations of cascaded and Thinker-Talker systems. The model demonstrates strong performance across various speech and audio tasks, including native full-duplex interaction, and introduces a method for intelligence-speaker decoupling to enable flexible voice customization.
Key insights
- A compact 7B-parameter LALM can achieve competitive performance across diverse tasks by co-designing architecture, multi-stage pre-training, and alignment-focused post-training.
- Decoupling dialogue intelligence from speaker characteristics in LALMs enables flexible voice customization with minimal TTS data, reducing deployment costs.
- Native full-duplex capabilities can be integrated into LALMs through dedicated pre-training, leading to more human-like, low-latency conversational interactions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.