正在加载视频...

视频加载失败

Efficient Streaming Audio-Visual Target Speaker Extraction for Real-World Acoustic Scenes paper:

16,262 次观看 • 20 天前 •via X (Twitter)

2 条评论

Hao 的头像
Hao19 天前

For video calls I'd check lip sync first. Studio's README lists about 0.7-0.8s of designed delay. Meeting video isn't delayed with the audio. The paper's 51ms is GPU inference.

Marvin 的头像
Marvin20 天前

RealSSA 把近场、远场和人数变化放进来,这个测试设置更贴近会议。还想看目标说话人转头或嘴部被遮挡时,提取效果会掉多少。

相关视频