Discord ↔ Gemini Live voice bot

A Node.js Discord bot that joins a voice channel, streams user speech to Gemini Live, and plays Gemini's streaming audio replies back into the same channel

What It Does

• Slash commands /join, /leave, and /reset
• Per-guild bridge session management, with one bridge per server
• Multi-user mixing so several people can talk to the bot at the same time
• Discord Opus receive -> PCM decode -> Gemini Live audio streaming
• Gemini Live PCM audio -> Opus encode -> Discord playback
• Tunable low-latency server-side VAD
• Local barge-in gating so background noise does not interrupt playback
• Pre-roll buffering so the start of a real interruption is preserved
• Optional root-level gem_sp.md support for the Gemini system prompt
• Automatic slash-command deployment on startup when the command JSON changes
• Bounded reconnect behavior on recoverable Gemini socket closes

Project Layout

src/
  app/
    commands.js       slash command definitions
    deploy/           application command sync
    events/           Discord event registration
    lifecycle/        process signals and graceful shutdown
  commands/           join/leave/reset handlers and command router
  config/             environment parsing and defaults
  services/
    audio/            PCM conversion, frame sizing, RMS, and mixing helpers
    bridge/           Discord-Gemini bridge orchestration and turn state
    gemini/           Gemini Live session lifecycle and reconnect handling
    voice/            receive mixer, playback queue, and voice lifecycle
  utils/              Discord messaging and preflight checks
  index.js            startup/bootstrap

Setup

cp .env.example .env
npm install
npm run start

On the first successful startup, the bot syncs /join, /leave, and /reset automatically.

Discord Permissions

The bot needs View Channel, Connect, and Speak in any voice channel it joins. Stage channels also need Request to Speak.

System Prompt Loading

If a file named gem_sp.md exists next to package.json, its contents are used as the Gemini system prompt. If not, the bot falls back to GEMINI_SYSTEM_PROMPT, then the built-in default prompt.

At startup the bot logs which prompt source it is using.

Environment Variables (.env)

This project reads config from src/config/index.js.

Required

• DISCORD_TOKEN - Discord bot token used for client.login(...).
• GEMINI_API_KEY - API key for @google/genai.
• GOOGLE_API_KEY is also accepted as a fallback name.

Optional

Variable	Default	Notes
GEMINI_MODEL	gemini-3.1-flash-live-preview	Gemini Live model passed to ai.live.connect(...).
GEMINI_VOICE_NAME	Pulcherrima	Voice used for Gemini speech output.
GEMINI_SYSTEM_PROMPT	built-in prompt	Fallback system prompt when gem_sp.md is absent.
ENABLE_SESSION_RESUMPTION	false	Enables Gemini session resumption handles across reconnects.
SUPPRESS_ROUTINE_ACTION_LOGS	false	Suppresses routine action logs (for example STT/TTS chatter and preflight diagnostics) while keeping warnings, errors, and important lifecycle events.
DISCORD_SPEECH_END_MS	350	Silence timeout for Discord receive streams. Lower values end streams faster; higher values tolerate longer pauses.
GEMINI_VAD_PREFIX_PADDING_MS	120	Server VAD pre-roll retained before detected speech starts.
GEMINI_VAD_SILENCE_DURATION_MS	350	Server VAD silence duration before Gemini ends speech input.
GEMINI_VAD_START_SENSITIVITY	START_SENSITIVITY_HIGH	Gemini start-of-speech sensitivity.
GEMINI_VAD_END_SENSITIVITY	END_SENSITIVITY_HIGH	Gemini end-of-speech sensitivity.
LOCAL_BARGE_IN_RMS_THRESHOLD	1700	Loudness threshold used before mixed live audio may interrupt Gemini.
LOCAL_BARGE_IN_CONSECUTIVE_FRAMES	3	Consecutive 20 ms frames required before interruption is accepted.
LOCAL_BARGE_IN_PREROLL_MS	240	Buffered live audio kept before a qualified interruption.
LOCAL_BARGE_IN_MIN_FORWARD_MS	450	Minimum live-audio forward window after interruption starts.
SERVER_INTERRUPT_FALLBACK_MS	1200	Releases the local interruption gate if Gemini's interruption acknowledgement is delayed.
GEMINI_CONNECT_TIMEOUT_MS	8000	Maximum time allowed for the Gemini Live connection to open.
GEMINI_SETUP_TIMEOUT_MS	12000	Maximum time allowed for Gemini setup to complete after connect.
GEMINI_RECONNECT_MAX_ATTEMPTS	6	Maximum consecutive reconnect attempts before the session enters degraded mode.
GEMINI_RECONNECT_BURST_WINDOW_MS	120000	Time window used to count reconnect bursts.
GEMINI_RECONNECT_STABLE_RESET_MS	120000	Stable period required before reconnect counters reset.
GEMINI_RECONNECT_BASE_DELAY_MS	1000	Base delay used for reconnect backoff.
GEMINI_RECONNECT_MAX_DELAY_MS	30000	Maximum reconnect backoff delay.
MIXER_SPEAKER_QUEUE_CAP_FRAMES	50	Max buffered 20 ms input frames kept per active speaker in the mixer.
PLAYBACK_PACKET_QUEUE_MAX_FRAMES	600	Max buffered outbound Opus frames before oldest playback frames are dropped.

Commands

• /join - connect the bot to your current voice channel
• /leave - disconnect the bot from the current guild's voice channel
• /reset - reset the current Gemini Live session or conversation state

Notes

• This build uses @discordjs/opus for Discord playback/input encoding instead of opusscript, which avoids the WebAssembly memory access out of bounds encoder failure seen in the playback path.
• The resampling and mixing path is intentionally simple and lightweight. It is suitable for a starter bot, not a studio-grade DSP chain.
• Gemini receives one mixed audio stream, not isolated speaker tracks, so speaker attribution in transcription is best-effort rather than guaranteed.
• The bot auto-syncs application commands on startup and caches the last deployed command hash in .cache/discord-commands.sha256.