@pipedog/audio-ducking

Ducking for pipedog: the bot's voice drops while the user talks over it, so it sounds like the bot heard them, and comes back when they stop, when the overlap becomes an interruption, or when the bot stops. Not in upstream Pipecat.

The output transport queues the bot's audio well before it plays, so the gain is applied where the audio leaves: AudioDuckingFilter is an output audio filter (audioOutFilter), which works with any transport or serializer. Its observer ducks it from the user's voice activity.

import { AudioDuckingFilter } from "@pipedog/audio-ducking";

const ducking = new AudioDuckingFilter({ duckedGain: 0.35 });
const params = new TransportParams({ audioOutEnabled: true, audioOutFilter: ducking });
const worker = new PipelineWorker(pipeline, { observers: [ducking.observer] });

• VADUserStartedSpeakingFrame ducks; VADUserStoppedSpeakingFrame, InterruptionFrame and BotStoppedSpeakingFrame restore. duck() and restore() can be called directly.
• The gain ramps over the first chunk after a change, so there is no click.
• FilterUpdateSettingsFrame({ settings: { duckedGain } }) changes the gain at runtime, and FilterEnableFrame turns ducking on and off.

Runtimes

| | Bun | Node ≥ 22 | Deno | Cloudflare Workers | |---|---|---|---|---| | AudioDuckingFilter | ✅ | ✅ | ✅ | ✅ |

Needs: nothing. Pure TypeScript.