Noise Cancelling Features with Self-Hosted Agents

Does this label mean that noise cancelling plugins like Krisp or ai-coustics will be affected in any way?

For the context - I have a project now where both plugins are running inside the agent process on Railway and authenticate through LiveKit Cloud. From testing, both appear to work correctly.

Is “self-hosted” here purely a deployment classification, or does it actually restrict any plugin functionality? In this case as far as I understand it’s still LiveKit Cloud hosted SFU rather than self-hosted, with only the agents logics being outsourced to a different place.

@Pawel_Lach, The “Self-hosted agent” label refers to your agent deployment location (Railway vs LK Cloud Agent Deployment), not the SFU. The features it gates are around managed agent deployment, observability/insights, and similar LK-Cloud-deployed-agent perks. Noise cancellation plugins aren’t in that gated set.

For Krisp and ai-coustics specifically, the docs note: “By default the ai-coustics plugin authenticates and meters usage through LiveKit Cloud” [ Noise & echo cancellation | LiveKit Documentation ]. That default auth path works for any agent that connects to LK Cloud, whether the agent runs on LK Cloud’s deployment infrastructure or on your own (Railway in your case). The only scenario where you’d switch to a self-supplied license key via the auth parameter is if you moved the SFU off LK Cloud, which the same page covers.

So your current setup (LK Cloud SFU + agent on Railway + both plugins) is on the supported default path. The “self-hosted agent” warning is true at the agent-deployment layer but doesn’t affect plugin functionality, which matches your empirical “both work correctly” observation.

In the meantime, for full self-hosted setups looking at OSS alternatives, RNNoise (Xiph, real-time, lightweight RNN) and DeepFilterNet 3 (Apache-2.0, deep-learning, real-time) are the two production-grade open paths.

I have an RNNoise variant trained on a large telephony-noise corpus that I’ll be open-sourcing soon, targeted specifically at telephony voice AI Krisp territory; happy to share when it lands.