RNNoise filter

OverviewCopy Anchor Link

RNNoiseFilter removes background noise from the user's audio before it reaches VAD and speech-to-text, using RNNoise, a small recurrent neural network. It is a port of Pipecat's RNNoiseFilter. Upstream calls the native pyrnnoise library; pipedog runs RNNoise 0.2 as WebAssembly, shipped inside the package, so there is nothing to install and it works on Cloudflare Workers too.

InstallationCopy Anchor Link

bun add @pipedog/audio-rnnoise-filter

import { RNNoiseFilter } from "@pipedog/audio-rnnoise-filter";

RuntimesCopy Anchor Link

Generated from pipecat.runtimes and pipecat.compat in packages/audio-rnnoise-filter/package.json: ✅ means the compat smoke test passes on that runtime.

UsageCopy Anchor Link

Set the filter as the transport's audioInFilter. It starts and stops with the transport.

import { RNNoiseFilter } from "@pipedog/audio-rnnoise-filter";
import { TransportParams } from "@pipedog/core/transports/base_transport";

const params = new TransportParams({
  audioInEnabled: true,
  audioInFilter: new RNNoiseFilter(),
});

RNNoise works on 480-sample frames (10 ms) at 48 kHz. The filter buffers incoming audio into frames and, when the transport is not at 48 kHz, resamples in and out with the soxr-style stream resampler (@pipedog/resampler-wasm). A call returns the frames that completed, so the first calls can be empty, and the output lags the input by less than one frame.

Send a FilterEnableFrame to switch the filter on and off while the bot runs:

import { FilterEnableFrame } from "@pipedog/core/frames/frames";

await worker.queueFrame(new FilterEnableFrame({ enable: false }));

ConfigurationCopy Anchor Link

new RNNoiseFilter(options?):

Lower-level pieces are exported too: initRnnoise(source?) loads the module from bytes, a WebAssembly.Module, a URL or a Response (on Workers only a module), RNNoise mirrors pyrnnoise's denoise_chunk, and RnnoiseEngine.createState() gives the raw 480-sample processFrame.

Cloudflare WorkersCopy Anchor Link

workerd refuses to compile WebAssembly from bytes, so on Workers the package's "workerd" export condition imports rnnoise.wasm as a pre-compiled module. Wrangler's default rule for .wasm files does this; nothing else is needed. The module is 1.4 MB.

Differences from upstreamCopy Anchor Link

• WebAssembly instead of pyrnnoise. The model is RNNoise 0.2, so the denoising is the same kind of output, but it is not bit-identical. The module is the one built by @jitsi/rnnoise-wasm (Apache-2.0).
• The rnnoise option and a protected _createResampler() hook replace the module-level patching upstream's tests do.

See the package README for the licenses of what it bundles.