LoopbackAudioInputDetector — loopback-audio-input

A new Pipeline Disruption detector on outbound audio tracks. It raises loopback-audio-input when the capture device’s label names a loopback of the machine’s own output (Monitor of … on PulseAudio/PipeWire, Stereo Mix / Wave Out Mix / What U Hear on Windows). The far end hears itself on such a call while every stat looks healthy. A label never changes, so each track is judged once and the detector then removes itself from the track’s registry. Config: loopbackAudioInputDetector.labelPatterns, defaulting to DEFAULT_LOOPBACK_AUDIO_INPUT_LABEL_PATTERNS. Track attribute: OutboundTrackMonitor.loopbackAudioInput.

AUDIO_INPUT_DEVICE client metadata

Each outbound audio track now sends one AUDIO_INPUT_DEVICE meta item when it is first monitored: { peerConnectionId, trackId, label, deviceId?, groupId? }. It reaches the server whichever source binding created the track, unlike MEDIA_TRACK_ADDED. ClientMetaTypes and AudioInputDevice are now exported.

New: session descriptions — SdpMonitor

ClientMonitor.acceptLocalDescription(peerConnectionId, description) and acceptRemoteDescription(...) hand an applied description to the peer connection’s new PeerConnectionMonitor.sdp (SdpMonitor, which has the same two methods). Each accepted description is added to the sample as LOCAL_SDP / new REMOTE_SDP metadata, payload { peerConnectionId, type, sdp } with a=ice-pwd redacted; identical repeats and rollbacks are ignored.

What it reads is published on PeerConnectionMonitor: negotiationRole, dtlsRole, remoteIceLite, localIceLite, bundled, negotiatedAudioCodecs, negotiatedVideoCodecs, sendingAudioDtx, receivingAudioDtx, sendingAudioInbandFec, receivingAudioInbandFec, audioRedNegotiated, sendingSimulcast; per media section on sdp.negotiatedMediaSections. The parser (parseSdp) is exported. No new dependency.

Sources feed it automatically: RtcPeerConnectionBinding and MediasoupTransportBinding read the connection’s descriptions when bound and on every signalingstatechange. The mediasoup binding reaches the connection through mediasoup-client’s private transport.handler._pc (peerConnectionOfTransport); a handler without one is monitored as before, without SDP.

Breaking: InventedSpeechDetector renamed to ConcealedSamplesDetector

“Concealed samples” is the term audio engineers recognise — it is what the W3C stats (concealedSamples, silentConcealedSamples) and NetEQ call it. “Invented speech” also overstated what the detector sees; see the next section. Straight rename, same algorithm, no aliases for any of the old names.

The issue payload gains concealmentEventRate, so a few longer gaps can be told from constant micro-concealment.

Why the old detector could not see dropouts

NetEQ fades concealment out: the mute slope steepens on the 3rd and 7th consecutive expand, and the output reaches zero after roughly 60–120 ms. From then on libwebrtc counts every concealed sample as silentConcealedSamples — the same bucket as DTX comfort noise, which the detector has to subtract. So each concealment event adds at most ~100 ms to the ratio however long the gap lasts, and a two-second dropout was nearly invisible. What the detector actually finds is frequent short gaps (choppy audio), and it is now documented as such.

New: AudioInterruptionDetector — audio-interruption

Reports dropouts of 150 ms and longer, from Chromium’s non-standard inbound-rtp interruptionCount / totalInterruptionDuration (libwebrtc counts every concealment event of at least 150 ms, silent part included). Same leaky-accumulator shape as its sibling: allowedInterruptedRatio: 0.02, raiseAfterInterruptedMs: 500. The interruption that ends when a paused stream resumes is discarded. Chromium only; elsewhere it reports inputsUnavailable. Config audioInterruptionDetector, monitor event audio-interruption, charged in DefaultScoreCalculator as audioInterruptionSeverity × AUDIO_INTERRUPTION_MAX_CHARGE (2).

New on InboundRtpMonitor: interruptionCount, totalInterruptionDuration, deltaInterruptionCount, deltaTotalInterruptionDurationInMs. They are not added to the sample schema.

Source: GitHub release 4.10.0. Published at 2026-10-04T16:53:02Z (UTC).