vmbridge

⚠️ WIP — Work In Progress. Functional but unfinished. not guaranteed to work on other environments than mines. Known unfinished item: clock-drift compensation (long sessions may develop clicks/dropouts). Use at your own risk.

Run Voicemod (Windows-only) on Linux by bridging its processed audio out of a Qemu VM and exposing it as a real, selectable Linux microphone. The virtual mic is registered on start and torn down on exit automatically — no manual pactl scripts.

Successor to the original proof-of-concept, archived in poc/.

How it works

 mic → Voicemod → VB-Cable In → "CABLE Output" → vmbridge-tx ──UDP──→ vmbridge-rx → virtual mic → Discord/OBS
        (Windows / QEMU guest)      WASAPI capture     raw PCM        (Linux host)   "Voicemod"

The QEMU guest and Linux host share a local virtual bridge (effectively localhost), so audio is sent as raw PCM — no codec, no encode/decode latency.

Project layout

cmd/
  vmbridge-rx/      Linux receiver + virtual-mic lifecycle
  vmbridge-tx/      Windows transmitter
internal/
  wire/             UDP packet format (12B header + s16le PCM)
  source/           virtual mic: load/unload module-pipe-source in-process
  receiver/         UDP ingest: gap-fill with silence + write to mic + stats
  capture/          WASAPI capture via malgo (Windows-only; stubbed elsewhere)
poc/                archived proof-of-concept (separate go.mod)

---

Prerequisites

Linux host

• Go 1.26+
• PipeWire or PulseAudio with pactl (pipewire-pulse or pulseaudio-utils)
• A working QEMU install with Windows running

Windows guest (inside QEMU)

• Voicemod installed
• VB-Cable installed
• Go for Windows
• A C toolchain — malgo uses cgo. Easiest: w64devkit (unzip, add …\w64devkit\bin to PATH). Verify with gcc --version.

---

Build & run

1. Linux host — receiver

go build -o bin/vmbridge-rx ./cmd/vmbridge-rx
./bin/vmbridge-rx

On start it registers a mic named Voicemod and listens on UDP :5000. On Ctrl-C (or SIGTERM) it unregisters the mic and cleans up. Leave it running the whole time you use Voicemod.

Open the UDP port if a firewall is active:

sudo firewall-cmd --add-port=5000/udp     # firewalld
# or
sudo ufw allow 5000/udp                    # ufw

Find the host IP the guest sends to:

ip -4 addr | grep inet | grep -v 127.0.0.1

Flags: -listen :5000 -name voicemod -desc Voicemod -rate 48000 -channels 2 -maxfill 50 -fifo <path>

2. Windows guest — transmitter

In PowerShell, inside the copied project folder:

go env -w CGO_ENABLED=1
go mod tidy
go build -o vmbridge-tx.exe ./cmd/vmbridge-tx

.\vmbridge-tx.exe -list                                     # find the VB-Cable capture device
.\vmbridge-tx.exe -addr <LINUX_IP>:5000 -device "CABLE Output"

undefined: malgo.* at build = cgo is off. Confirm gcc --version works and go env CGO_ENABLED prints 1.

Flags: -addr host:port -device "CABLE Output" -rate 48000 -channels 2 -frame 5 -queue 8 -list

3. Voicemod routing (Windows)

1. Voicemod → Settings → Output = CABLE Input (VB-Audio Virtual Cable).
2. Enable monitoring ("Hear myself").
3. Voice Changer ON, pick a voice.

4. Use it (Linux)

In Discord/OBS, select input Voicemod.

• Apps enumerate devices at launch — start vmbridge-rx first, then (re)start the app. Discord: fully Quit from tray, not just close the window.

Live-monitor your own changed voice (headset only — speakers cause feedback):

pactl load-module module-loopback source=voicemod sink=@DEFAULT_SINK@ latency_msec=20

---

Verify

vmbridge-rx logs rx pkts= every 5s — a rising count means audio is flowing. Or capture directly from the virtual mic:

pw-record --target voicemod /tmp/t.wav   # talk, Ctrl-C
pw-play /tmp/t.wav

Troubleshooting

Symptom	Fix
rx pkts=0	Wrong -addr IP or UDP 5000 blocked. Check firewall.
package cmd/... is not in std	Use ./cmd/... (with leading ./).
undefined: malgo.* (Windows)	cgo off — install gcc, go env -w CGO_ENABLED=1.
Mic missing in Discord/OBS	Start rx first, then fully restart the app.
no device matching (Windows)	Run -list, copy the exact substring into -device.
Choppy after several minutes	Clock drift — known WIP gap, not yet fixed.
Robotic/garbled	Sample-rate mismatch — keep 48000 everywhere.

Known limitations (WIP)

• No clock-drift compensation — independent capture/playback clocks diverge over time → eventual clicks/dropouts on long sessions. Planned: adaptive resampling driven by FIFO fill level.
• Native libpipewire source node (lower latency than the FIFO) not yet done.
• Mic into the VM is assumed handled by QEMU's own passthrough.