plat 122

← Pulse platform issue index

PLAT-122 — real microphone reads digital silence through every app on a dev machine

Coordination	Value
Assigned agent	unassigned
Ticket state	open — environment-level, not yet root-caused
Last synchronized	2026-08-16

• Priority: P1 — blocks live verification of PLAT-120 (Video Studio voice dictation); not reachable from workflow code, external action required.
• Owner: none (OS/hardware/driver layer) — see voice_dictation_mic_captures_silence.md for the full investigation this ticket indexes.

Problem

On the reporter's dev machine, agent_go/pkg/voicestt's server-side RMS telemetry (pcmRMS in cmd/server/voice_stt_routes.go) shows every audio chunk at rms=0.0000 during real mic dictation — genuine digital silence reaching the backend, not a transport or model defect.

What was ruled out (full detail in the linked doc)

• The STT pipeline itself: proven correct via a raw WAV file streamed through the live WebSocket endpoint (correct transcripts), and a synthetic 440Hz tone fed through the actual browser capture code (real, varying PCM samples, decoded RMS ≈ 0.71).
• Wrong input device: a bare getUserMedia call in the reporter's own tab was granted the correct device (MacBook Air Microphone (Built-in)), returned muted=false enabled=true, and still read maxPeak=0 over 2+ seconds of real speech.
• Browser-automation tooling serving a fake device: reproduced in a second browser window not under any automation control.
• SparkQuill's Electron app holding the input device: reproduced with it fully quit (Cmd+Q).
• A third-party noise-suppression HAL driver (e.g. Krisp): none present; /Library/Audio/Plug-Ins/HAL/ has only Apple's own ParrotAudioPlugin and the (unused-here) Zoom virtual device.
• macOS Privacy & Security → Microphone: Chrome is enabled there.

Remaining candidate, unconfirmed

macOS System Settings → Sound → Input has a level meter and a separate Input volume slider. The meter can reflect raw hardware signal presence even when the volume slider — the actual gain handed to apps via CoreAudio — is near zero, which would explain real signal at the hardware level and exact silence in every application. Asked the reporter to check; unconfirmed as of this ticket.

Acceptance

• Either the Input-volume slider is confirmed as the cause and the ticket closes with that as the documented resolution (a config/environment fix, not a code change), or
• a fresh elimination pass (see the linked doc's "Follow-up if this recurs" section) locates a real code-owned cause, and this ticket re-opens against the owning component instead of environment.