Free, headless, bidirectional real-time voice pipe over a local network.
vox runs as two symmetric instances, one per machine. Each captures audio from a local device, Opus-encodes it, and sends it over plain UDP to the other; at the same time it receives, decodes, and plays the other side's audio. There is no "server" and no "client" — just two peers that know each other's address.
It was built to solve one specific gap: GPU desktop-streaming hosts like Apollo (a Sunshine fork) with Moonlight carry video and speaker audio from the host, but have no microphone backchannel. vox supplies that missing duplex — and works as a general-purpose LAN voice pipe for anything else.
The live --output tui dashboard: device & peer status, throughput, the adaptive
jitter buffer, and live loss / recenter / drift readouts.
- Free & FOSS. No paid tiers, no accounts.
- Headless. Fully scriptable from arguments or a config file. No GUI.
- Low latency. Tuned for real-time voice (Teams, Discord, in-game comms).
- Cross-platform. Windows ↔ Linux and Windows ↔ Windows.
- No extra audio servers. Uses what's already there — WASAPI on Windows, ALSA/PipeWire on Linux. No JACK, no Steam, no intermediary daemon.
Each instance runs four threads: a capture callback and a playback callback
(audio I/O), plus a send thread (encode → UDP) and a receive thread (UDP →
decode). A lock-free ring buffer hands audio off the real-time callbacks, and a
jitter buffer on the receive side smooths out network timing and adapts its depth to
the measured jitter. Opus runs at 48 kHz, 20 ms frames, mono; in-band FEC for
graceful packet loss is available (opt-in via --fec, off by default — on a clean
link it costs more than it saves).
See docs/DESIGN.md for the full architecture and rationale.
Config-file mode (used by Apollo's command hooks):
vox config.toml
Ad-hoc mode:
vox --peer <host[:port]> [--bind <port>] [--capture <dev>] [--playback <dev>]
--capture / --playback name the local device to record from / play to.
Accepts default, none (disable that direction), or a device name. Omitted
means default. Port defaults to 9680 when omitted.
Run vox --list-devices to see exact device names (--capture/--playback
match them exactly). Other handy flags: --output tui for a live dashboard
(throughput, loss, jitter), --output quiet for silent headless runs, -v for
technical logs, and --print-config to dump the resolved settings.
Example — Windows host (VB-Cable A = desktop audio, B = virtual mic):
vox --peer 192.168.1.20 --capture "CABLE-A Output" --playback "CABLE-B Input"
Example — client (real mic and headphones):
vox --peer 192.168.1.10
Audio tuning (--bitrate, --fec, --jitter-ms, --drift-correct, per-device
sample rate, etc.) lives in the TOML config and can be overridden by flags. See
docs/DESIGN.md §6–§7.
Drift correction. If the TUI's drift readout sits at a steady nonzero value
(the jitter buffer slowly filling or draining over minutes — independent clocks on
the two machines), enable --drift-correct: it smoothly holds the buffer at its
target, removing the occasional clock-drift cutoff and keeping latency low. It's off
by default because it only helps where drift exists — on a clean link it just adds a
resampler for no gain.
Rust + Cargo. Requires a C compiler and cmake (to build bundled libopus); on
Linux also alsa-lib-devel.
cargo build --release
MVP builds natively per target (build Windows on Windows, Linux on Linux). Cross-compiling from Fedora is a later-phase convenience.
Working beta (pre-1.0; no public release cut yet). Phase 1 MVP and most of Phase 2 are done: full-duplex LAN
voice with a clap CLI + TOML config and a live TUI, plus in-band FEC + graceful loss,
reconnection robustness, non-48k device resampling, and an adaptive jitter buffer —
verified machine-to-machine. Remaining for Phase 2: smooth clock-drift compensation
and the Linux client; Phase 3 covers packaging, encryption, and an Android front-end.
Not a 1.0 release yet. See docs/PLAN.md for milestones.
In: two-peer duplex voice over LAN, device selection, adjustable quality, graceful loss. Planned for Phase 3: payload encryption. Out: more than two peers, GUI, audio mixing/effects.