← sources

AudioGhost v1.2: remove any sound from audio or video by typing what it is (local, open source, built on Meta's SAM-Audio)

source: reddit (reddit:LocalLLaMA)published 2026-11-18 05:17ingested 2026-11-19 02:39

retained text

AudioGhost randomly got a wave of attention this week (100 stars in a day, no idea where from), so I took that as a sign to finally ship the update I'd been sitting on.                  

Quick recap if you haven't seen it: you drop in an audio or video file, type the sound you're after ("background music", "dog barking", "people talking"), and it pulls that sound out or removes it. It's a local web app around Meta's SAM-Audio, so everything runs on your own GPU and nothing gets uploaded anywhere.

The use case I hear about most: you film a vlog or an interview outside, a shop nearby is playing music, and YouTube flags the video. Now you can type "background music", mute that track and export the video with everything else still there.

**What's new in v1.2:**  
\- Several sounds per run, each on its own track, with mute / solo / volume like a tiny DAW  
\- Export exactly what you hear (WAV / FLAC / MP3), or the video with the new soundtrack  
\- The "everything else" track is now generated by the model instead of subtracted, so removing a sound works a lot better  
\- Mark on the timeline where a sound is (or isn't) to point the model at the right thing  
\- New UI, and your sessions survive reloads and restarts

If you tried an old version: it's also 4-6x faster than the first release.

Hardware: the small model runs in a bit over 4 GB of VRAM, so most NVIDIA cards are fine. Windows and Linux have install scripts. macOS works too, but it runs on the CPU for now, so it's slow.

One caveat: the official SAM-Audio pipeline also loads several helper models (ImageBind, CLAP, and a judge model that picks the best of several outputs). They add about 25 GB of downloads and need far more VRAM. I stripped those out so it fits on normal GPUs, which means quality is a bit below the full pipeline and results vary from run to run. If a track comes out empty or rough, delete it and run it again.

**Demo video and install instructions:** [**https://github.com/0x0funky/audioghost-ai**](https://github.com/0x0funky/audioghost-ai)

I'd love feedback, especially from anyone who tries it on messy real-world recordings. Bug reports are welcome on GitHub.

document fetched 2026-11-19 00:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1xz6xl7/audioghost_v12_remove_any_sound_from_audio_or