Audio-Pipeline-Plugin

• A WebAssembly module implementing fixed-order audio pipeline processing for web frontends.
• The denoise stage currently supports rnnoise and deepfilternet.

Denoise Stage Modules

• rnnoise (default)
• deepfilternet

DeepFilter Build Artifacts

• DeepFilter wasm source package: DeepFilterNet/libDF/pkg/df.js + df_bg.wasm.
• Artifact generator: node ./scripts/build-deepfilter-artifacts.mjs.
• Generated files: src/dist/deepfilter-bindgen.js, src/dist/rnnoise-sync.js, dist/deepfilter.wasm.
• build:js runs artifact generation first, then bundles worklet + library outputs.

Example

import { AudioPipelineTrackProcessor } from "@cc-livekit/audio-pipeline-plugin"

const processor = new AudioPipelineTrackProcessor({
    workletUrl: "/assets/AudioPipelineWorklet.js",
    stages: {
        denoise: "rnnoise",
    },
    moduleConfigs: {
        rnnoise: {
            vadLogs: true,
            vadLogIntervalMs: 1000,
        },
        deepfilternet: {
            attenLimDb: 100,
            postFilterBeta: 0,
        },
    },
})

await processor.setModuleConfig("rnnoise", { vadLogs: true, vadLogIntervalMs: 800 })
await processor.setModuleConfig("deepfilternet", { attenLimDb: 30, postFilterBeta: 0.02 })
await processor.setStageModule("denoise", "deepfilternet")
await processor.setStageModule("denoise", "rnnoise")

Runtime Behavior

• Single AudioWorklet runtime.
• WASM and models are loaded only at init. Both rnnoise and deepfilternet WASM binaries are fetched and both modules are created during init(). No model or WASM is loaded when switching modules or when calling setModuleConfig later.
• setStageModule("denoise", ...) switches the active denoise module in-place (rnnoise ↔ deepfilternet). Only the already-loaded active module is used; no extra loading.
• setModuleConfig("deepfilternet", ...) after init only updates parameters (attenLimDb, postFilterBeta). DeepFilter uses the built-in model only (no custom model loading).
• setModuleConfig("rnnoise", ...) updates rnnoise-only config.
• vadLogs and vadLogIntervalMs are rnnoise-only configs.
  • They are ignored by deepfilternet.
  • Updating rnnoise config while deepfilter is active is cached and applied when switching back to rnnoise.
• Input audio must be 48 kHz; the pipeline does not perform sample-rate conversion.