Voice Gateway

Omni ships a Bun-native voice gateway that joins Discord voice channels, decrypts incoming audio, and exposes the streams over a local WebSocket so agents can listen, transcribe, and respond. The implementation lives in @omni/voice-client and is wired into the Discord channel adapter — no external Node sidecar required.

What it does

Capability	Notes
Discord voice gateway v8	UDP transport, libsodium-backed packet auth
DAVE E2EE	Discord Audio & Video End-to-End encryption — Omni participates as a full DAVE client (group key exchange, sender keys, key rotation)
Opus codec	Decode incoming Opus frames, optionally re-encode for outbound
Per-user streams	Audio is demuxed per Discord user ID so transcription/STT can target a single speaker
Session WebSocket	A local WebSocket (`omni voice stream`) emits audio frames + control events for downstream consumers

Lifecycle

omni voice join ──▶ Discord gateway handshake
                         │
                         ▼
                    DAVE key exchange
                         │
                         ▼
              Voice UDP / SRTP established
                         │
                         ▼
           Per-user Opus streams demuxed
                         │
                         ▼
        omni voice stream <id> taps the bus
                         │
                         ▼
                 omni voice leave

CLI

# Join a Discord voice channel
omni voice join --instance my-discord --guild <guild-id> --channel <voice-channel-id>

# List active voice sessions
omni voice sessions

# Tap a session over WebSocket — opus by default, pcm for raw frames
omni voice stream <session-id> --format pcm --save ./recordings

# Filter to a single speaker
omni voice stream <session-id> --user <discord-user-id>

# Show only control events (joined/left/speaking) without audio stats
omni voice stream <session-id> --events-only

# Leave the session
omni voice leave --instance my-discord

omni voice stream exits cleanly on Ctrl+C and is safe to pipe into transcription tooling.

Discord-specific notes

The Discord adapter manages voice per instance — one Discord bot token, one voice gateway. Multiple guild channels can be joined sequentially but only one at a time per instance today.
DAVE is mandatory for any guild that has it enabled. The voice client negotiates the protocol version automatically and rejects sessions that fail the handshake rather than falling back to plaintext.
Audio frames are never persisted by the gateway itself. Use --save <dir> on omni voice stream if you need on-disk artefacts; otherwise the bus is in-memory only.

DAVE rejection is loud — when a session fails handshake the gateway emits a voice.handshake_failed event. Check omni events list --type voice.handshake_failed --since 1h if voice goes dark in a DAVE-enabled guild.

Voice CLI verbs

omni voice join, leave, stream and the speak/listen verbs.

Instances

Discord bot token configuration and per-instance setup.

Media architecture

Media storage, transcription, and batch backfills.

Events

Stream voice.* events for live monitoring.

Getting Started

Core Concepts

CLI Reference

Architecture

Voice Gateway

Voice Gateway

What it does

Lifecycle

CLI

Discord-specific notes

See also

Voice CLI verbs

Instances

Media architecture

Events

​Voice Gateway

​What it does

​Lifecycle

​CLI

​Discord-specific notes

​See also

Voice CLI verbs

Instances

Media architecture

Events

Voice Gateway

What it does

Lifecycle

CLI

Discord-specific notes

See also