audio_flutter_linux #

Linux implementation of the audio_flutter platform contract: microphone and system-audio capture, plus PCM playback, over the PulseAudio/PipeWire command-line tools.

The package contains no native code. Capture spawns parecord and falls back to pw-record; playback spawns paplay and falls back to pw-play. Both are asked for signed 16-bit little-endian PCM at the requested rate and channel count, which this package converts to and from the float32 frames the contract carries. Registration is dartPluginClass only — there is no C++ registrant to build.

Targeting #

PulseAudio exposes one flat source namespace: hardware inputs (alsa_input.*) and sink monitors (alsa_output.*.monitor) are both sources addressable by name. PlatformCaptureRequest.inputDeviceId therefore carries a PulseAudio source name for both capture kinds:

No new request field was added: a separate sourceId would duplicate inputDeviceId on a platform where both kinds address the same namespace, and it would collide in name with PlatformCaptureSessionInfo.sourceId, which already means something else.

listAudioInputDevices() returns the non-monitor sources. listSystemAudioSources() — a Linux-only addition, not part of the platform contract — returns the monitors, so a UI can offer them for selection. PlatformCaptureRequest.processIds has no Linux equivalent and is ignored; listAudioProcesses() is always empty, because a monitor mixes a whole sink and there is nothing analogous to a Core Audio per-process tap.

Permissions #

Linux has no system-audio permission gate: monitor sources are readable by any client of the running sound server. isSystemAudioCaptureSupported() reports whether a capture tool and pactl are installed, and requestSystemAudioCapturePermission() mirrors it. Both answer "is the mechanism present", never "did a user grant something".

Format support #

Formats map straight onto process arguments (--rate, --channels), so any rate up to 192 kHz and up to 32 channels is accepted. Combinations PulseAudio cannot express — or a frame duration too short to hold one whole sample frame — are rejected from prepareCapture/preparePlayback with a LinuxAudioFormatException rather than failing mid-capture.

Health #

captureEvents reports starting, then running with receivingAudio: true once bytes actually arrive. A source that starts cleanly but delivers nothing fails after two seconds with code SystemCaptureDead, matching the Darwin watchdog. An unexpected tool exit fails with CaptureProcessExited and an excerpt of the child's stderr. The child's stderr is always drained, because an unread pipe fills its kernel buffer and blocks the process.

Backpressure #

readCaptureFrames pulls from a bounded ring sized by maxBufferedDuration ÷ frameDuration, honouring every PlatformCaptureOverflowPolicy. Frame sequence and sampleOffset are assigned as audio is produced, so they keep advancing across drops and the frame delivered after a gap reports droppedFramesBefore. Playback backpressure is the child's stdin: every write awaits its flush.

Status #

Not yet validated against a real PulseAudio or PipeWire server. Every command line, parse, and lifecycle path is unit-tested through an injected process seam, and that suite runs on any host — but no audio has been captured or played on Linux hardware. Treat argument compatibility across PulseAudio and PipeWire versions, monitor availability, and end-to-end latency as unverified until that run happens.

Attribution #

Command assembly, pactl parsing, monitor resolution, the capture-tool fallback order, and the always-drain-stderr rule are derived from Control Center, MIT © 2026 Samuel Alev. See NOTICE.