@proj-airi/audio owns browser audio resources and format conversion. @proj-airi/pipelines-audio owns capture intervals, observation windows, and playback groups.

Use it for

Open a microphone, a borrowed MediaStream, or an audio file as a PCM source.
Encode PCM as WAV, or convert PCM to provider input formats.
Play files or streamed PCM through Web Audio.
Fade owned playback nodes before reporting silence.

Audio sources

A source is an AudioSource. Its open(signal) call returns a PCM stream. Aborting the signal releases the connection.

Source	Module	Live	Owner of the tracks or data
`microphoneSource(constraints, options)`	`@proj-airi/audio/browser`	Yes	The source. It stops the tracks and closes its AudioContext.
`mediaStreamSource(stream, context)`	`@proj-airi/audio/browser`	Yes	The caller. The source disconnects only its own nodes.
`fileSource(blob)`	`@proj-airi/audio/encoding`	No	The caller. Mediabunny decodes the file for each connection.

A live source cannot wait for its readers. Wrap a live source in an AudioInput to share it. Each consumer subscribes with its own signal.

A file is not live. Its reader sets the decode speed. Each consumer opens its own stream:

const events = transcriber.transcribe({ audio: fileSource(recording).open(signal), signal })

import { microphoneSource } from '@proj-airi/audio/browser'
import { encodeWav } from '@proj-airi/audio/encoding'
import { AudioInput, capture } from '@proj-airi/pipelines-audio'

const input = new AudioInput(microphoneSource({ audio: { echoCancellation: true } }), { historyMs: 360 })

// Start this in the click handler. getUserMedia starts before the first await.
const recording = capture(input)
const wav = encodeWav(recording.stream, { sampleRate: 16000, channels: 1 })

// Call finish when the control is released. Call cancel to discard the interval.
await recording.finish()
showPreview(await wav)

The first subscriber opens the microphone. The last subscriber to leave stops the tracks. A late permission result after that point stops its tracks immediately.

toMediaStream(stream, context, signal) plays PCM into new tracks. Use it for APIs that accept only a MediaStream, such as Web Speech recognition.

Playback

import { BrowserPlayback } from '@proj-airi/audio/browser'
import { Playback } from '@proj-airi/pipelines-audio'

const playback = new Playback(new BrowserPlayback(audioContext))
const group = playback.openGroup('answer')
void group.enqueue({ id: 'first', audio: audioBlob })
const receipt = await group.stop({ fadeMs: 100 })

The stop receipt follows the audio clock and source completion. It does not rely on a UI timer. BrowserPlayback uses audioContext.currentTime as the playback clock. startAtMs and the receipt interval are in milliseconds on that clock. The caller owns audioContext and closes it after its consumers finish.

Do not use it for

Character routing, chat persistence, or agent notifications.
VAD policy, wake-word matching, memory search, or transcript rewriting.
Vue state or cross-window commands.

Checks

Run pnpm -F @proj-airi/audio typecheck and its Vitest projects. Node tests cover WAV encoding and file decoding. Browser tests cover Web Audio, microphone sharing, late permission, and track cleanup.