AI-based live audio content filtering for webOS TVs
Mute an offensive word in live TV audio before the viewer hears it, on the TV itself.
- real-time audio, on the TV
- Live
- edge models, no cloud
- On-device
- LG's TV platform
- webOS
Context
Viewers want broadcasts without offensive language, but muting has to happen live, on the TV, with no time to send audio anywhere else.
That rules out large server models: whatever listens has to run inside the TV's media stack under tight latency and compute limits, on noisy broadcast audio.
What I built
- A real-time pipeline that connects streaming speech analysis to the TV media stack.
- Changes to GStreamer-based playback pipelines and webOS media services so audio can be analysed and muted in the playback path.
- An evaluation of lightweight edge AI models under on-device limits, focused on low-latency inference and robustness to broadcast noise.
Outcome
- Found that speech recognisers trained on clean read speech systematically under-recognise profanity, which is the exact words the system needs to catch.
- After the product track was discontinued, that finding became independent research: the BLEEP dataset.