Voice dictation: "No audio detected from microphone" despite fully working mic (Arch Linux, PulseAudio)

Environment

• Claude Code version: 2.1.215 (native install, linux-x64, commit 316ce99628e8)
• OS: Arch Linux, kernel 7.0.12-arch1-1
• Audio server: PulseAudio (not PipeWire)
• sox present: SoX_ng v14.8.0.1 (the SoX Next Generation fork, API/CLI-compatible with classic SoX)
• arecord present (alsa-utils)
• Mic: USB audio interface (Soundpresence HD-II), correctly set as PulseAudio default source

Steps to reproduce

1. Set ~/.claude/settings.json → "voice": {"enabled": true, "mode": "tap"} (also reproduced in hold mode)
2. Start a Claude Code session, run /voice tap
3. With input empty, tap space to start recording, speak normally, tap again to stop
4. Result: error "No audio detected from microphone"
5. Fully quit Claude Code and started a brand-new session/terminal (to rule out a stale device handle from before the mic was ready) — same error persists

What I verified about the audio stack (all working correctly)

• pactl get-default-source → correctly resolves to the USB mic (alsa_input.usb-Soundprese_HD-II-00.analog-stereo)
• Mic not muted; PulseAudio volume 66%; ALSA hardware capture gain 77/99 (amixer -c 1 contents)
• sox -d file.wav trim 0 3 while speaking → clear signal captured: amplitude jumped from ~0.03 (ambient noise) to ~0.115 (speech), RMS 0.026
• timeout 4 rec -q --buffer 1024 -t raw -r 16000 -e signed -b 16 -c 1 file.raw (16kHz mono signed PCM — the format voice dictation likely expects internally) while speaking → also captured clear signal: amplitude 0.136, RMS 0.034
• Both sox/rec and arecord binaries are present and on PATH
• claude doctor reports "No installation issues found"

Since raw audio capture succeeds through every external tool, backend (-t pulseaudio, default, raw ALSA-via-pulse) and sample format tested (48kHz stereo and 16kHz mono), the mic/PulseAudio/ALSA configuration itself is not the problem. The failure appears to be specific to Claude Code's own audio capture path (built-in native module, with sox/arecord as documented fallback).

Expected behavior

Voice dictation should successfully capture and transcribe speech, given the microphone is fully functional and produces a clear signal via every external capture method tested.

Additional notes / requests

• Could not find any documented way to force the sox/arecord fallback instead of the built-in native audio capture module — this would help narrow down which path is failing.
• Could not find any voice-dictation-specific debug/verbose logging flag (checked for env vars like DEBUG, ANTHROPIC_LOG, etc.) to see which backend is actually being invoked and why it reports no audio.

Happy to run additional diagnostics or a debug build if one exists.