Enhanced Noise Cancellation Plugin for LiveKit

Add realtime enhanced noise cancellation to inbound AudioStream. Fully compatible with LiveKit Agents.

Audio chain

This package runs server-side on inbound audio (e.g. in an agent or on an AudioStream). Remote participants’ tracks are cleaned before your code or the agent consumes the audio.

flowchart LR
  Room["LiveKit room"] --> Track["Remote track (inbound)"]
  Track --> NC["livekit.plugins.noise_cancellation"]
  NC --> Stream["AudioStream / Agent input"]
  Stream --> Consumer["Your code or agent"]

Usage

In LiveKit Agents

Include the filter in RoomInputOptions when starting your AgentSession:

from livekit.plugins import noise_cancellation

# ...
await session.start(
    # ...,
    room_input_options=room_io.RoomInputOptions(
        noise_cancellation=noise_cancellation.BVC(),
    ),
)
# ...

On AudioStream

Noise cancellation can also be applied to any individual inbound AudioStream:

from livekit.rtc import AudioStream
from livekit.plugins import noise_cancellation

stream = AudioStream.from_track(
    track=track,
    noise_cancellation=noise_cancellation.BVC(),
)

Available Models

There are three noise cancellation models available:

# Standard enhanced noise cancellation (NC)
noise_cancellation.NC()

# Background voice cancellation (NC + removes non-primary voices
# that would confuse transcription or turn detection)
noise_cancellation.BVC()

# Background voice cancellation optimized for telephony applications
noise_cancellation.BVCTelephony()

Notes

Noise cancellation only needs to be applied once, so if you choose to apply it here you should disable noise cancellation / Krisp filter in your frontend clients.

If you experience crashes when using noise_cancellation (especially on AMD CPUs), it may be due to a failure in OpenBLAS's CPU detection. Manually setting the OPENBLAS_CORETYPE environment variable to a more conservative value (e.g., Haswell) may resolve the issue.