Murf BVC for LiveKit Agents

Background voice cancellation for voice agents. Murf BVC keeps the person talking to your agent and removes everyone and everything else (TV, other people in the room, street and office noise) before your agent's VAD, turn detection and speech-to-text see the audio. Fewer false interruptions, fewer transcripts of someone else's words.

It runs on CPU inside your agent process, like LiveKit's own noise-cancellation plugin. No GPU and no extra services; audio never leaves your process.

Requirements

• Python 3.10+ and livekit-agents 1.2+
• Linux (glibc 2.28+: Ubuntu 20.04+, Debian 10+, RHEL/Rocky/Alma 8+, Amazon Linux 2023) on x86_64 or aarch64, or macOS 11+ (Apple Silicon) / 10.15+ (Intel)
• x86_64 CPUs need AVX2 and FMA (Intel Haswell / 2013 or newer, any AMD Zen); on ARM, AWS Graviton2 or newer, or any Apple Silicon Mac
• A Murf API key with BVC enabled, or an offline license key for on-premise and air-gapped deployments (contact Murf)

Install

pip install livekit-plugins-murf-bvc

To test the 0.3.0 pre-release, pin it: plain pip install skips pre-releases and installs 0.2.2.

pip install --pre livekit-plugins-murf-bvc==0.3.0rc2

Both versions work against Murf. 0.3.0 seals the model key to each agent process and signs its usage reports.

Quick start

Set MURF_API_KEY (the same Murf API key the Murf TTS plugin uses; in .env.local like your other keys), call murf_bvc.prewarm() in the server's setup function, and pass murf_bvc.BVC() as the noise cancellation for the room's audio input:

from livekit.agents import Agent, AgentServer, AgentSession, JobContext, JobProcess, cli, room_io
from livekit.plugins import murf_bvc

server = AgentServer()


def prewarm(proc: JobProcess):
    murf_bvc.prewarm()  # license + models, before any call is assigned


server.setup_fnc = prewarm


@server.rtc_session()
async def entrypoint(ctx: JobContext):
    session = AgentSession(...)  # your STT, LLM, TTS, VAD, turn detection
    await session.start(
        room=ctx.room,
        agent=Agent(instructions="You are a helpful assistant."),
        room_options=room_io.RoomOptions(
            audio_input=room_io.AudioInputOptions(noise_cancellation=murf_bvc.BVC()),
        ),
    )


if __name__ == "__main__":
    cli.run_app(server)

That's all most agents need. The defaults are what we recommend for production. On-premise or air-gapped, set MURF_BVC_LICENSE="mbvc1...." (an offline license from Murf) instead of MURF_API_KEY.

Why prewarm

BVC() never blocks your agent. Without prewarm(), though, the first call in a process waits for work that prewarm() does up front: fetching the first lease for an API key (a request to Murf), and decrypting and loading the models (about 0.3 s, which would otherwise happen inside LiveKit's audio pipeline when the first participant's audio arrives). Until the lease arrives, audio passes through unprocessed.

LiveKit runs the setup function once per worker process while the process waits in the idle pool, so the time it takes is never on a call. It is the same place agents load their VAD model. prewarm() takes the same arguments as BVC(), returns True when ready, and returns False (logging why) if there is no valid license within timeout seconds (default 5); it never raises.

Keys

Like LiveKit's other plugins, BVC reads its key from the environment unless you pass one:

If you'd rather not put the key in the environment, pass it in code to both calls, for example murf_bvc.prewarm(api_key=key) and murf_bvc.BVC(api_key=key) with the key from your secrets manager. A key passed in code applies only to that BVC value and the streams it creates (a BVC() without a key is not licensed by it), never enters LiveKit's options, and is not inherited by subprocesses or visible in the process environment, where crash reporters and debugging tools often collect it.

Any code running in your agent process (other plugins included) can read memory it shares with the agent, so no way of passing a key hides it from them. Treat the BVC key like the agent's other API keys.

Options: when to change them

murf_bvc.BVC(api_key=None, license_key=None, restore=True, normalize=True, native=True)

Rooms with several participants

Nothing to change: one BVC() value serves every participant, each with its own independent state.

If you use native=False (or BVCProcessor directly), each processor handles one track, so pass a selector instead:

noise_cancellation=lambda params: murf_bvc.BVC(native=False)

What to expect

• Latency: 10 ms at 16 kHz, about 13.6 ms at 24 or 48 kHz (resampling included).
• CPU: about 3% of one core per active participant on an AWS AMD EPYC instance, plus a speaker check every 0.5 s of speech while caller restore is on. Processing is single-threaded per stream, so it scales with the number of concurrent calls.
• Audio: any sample rate LiveKit delivers (8–96 kHz); stereo input is mixed to mono.

License behaviour

Without a valid license BVC does not stop your agent: it logs an error and passes audio through unchanged, delayed by the same few milliseconds, so calls keep working and a license expiring mid-call never knocks audio out of sync.

With an API key, the first lease is fetched in the background (or by prewarm()); calls that start before it arrives pass audio through, then are processed once it does, without reconnecting. If Murf cannot be reached, processing continues on the current lease for at least 45 minutes; after it expires audio passes through until a renewal succeeds, then processing resumes on its own.

A refused key (HTTP 401, 402 or 403) is retried every 5 minutes, and audio passes through meanwhile. Once the account is fixed, processing starts within about 5 minutes without restarting the agent: for example after BVC is enabled on the account (it can take a few minutes to take effect) or credit is topped up. Restart the workers to pick it up immediately. A key passed in code or in MURF_API_KEY is read when the agent starts, so replacing the key itself needs a restart.

0.2.2 and earlier stop requesting leases after a refusal, so audio stays unprocessed until the workers restart.

Check a key at startup with the result of prewarm():

if not murf_bvc.prewarm():   # the reason is logged
    ...

Troubleshooting

All messages, including those from the native library inside LiveKit's audio pipeline, go to the Python logger livekit.plugins.murf_bvc, so they follow your agent's log level and format (JSON in production). License events carry license_sub and license_expires_at as structured fields.