140 upvotes
New streaming model pulls out target voice from noisy video
What is Efficient Streaming Audio-Visual Target Speaker Extraction for Real-World Acoustic Scenes?
The system listens to a video feed and isolates the speech of a chosen person in real time, even when other speakers, echo, or sudden noises are present. It runs efficiently enough for latency-sensitive apps like live calls or AR glasses.
Who it is for
Anyone building live transcription, video chat, or AR audio filters can use it to deliver clearer, speaker-specific audio without lag.
How to try it
Open the demo on the Hugging Face Space linked from the paper page