Voice

discli’s voice stack lets your bot join a server voice channel, transcribe everyone live, speak via TTS, and play audio. By the end of this guide you’ll have a meeting transcriber that writes speaker-labelled lines to a file and asks Claude to summarise the call when you stop it.

What you’ll build

A bot that joins a voice channel, prints lines like:

- **[14:32:08] Roy:** so what's our plan for the migration
- **[14:32:14] Sara:** I'd start with the read path first
- **[14:32:21] Roy:** good idea — can we ship that this sprint?

…to console and to ~/.discli/transcripts/meeting-<timestamp>.md, then on Ctrl+C asks Claude to produce a structured summary.

Step 1 — Install the voice extras

The core discord-cli-agent install is text-only on purpose — the voice stack pulls in PyNaCl, discord-ext-voice-recv, davey (Discord’s end-to-end voice encryption library), and audioop-lts. Install the extras when you want voice:

pip install 'discord-cli-agent[voice,deepgram]'

uv add 'discord-cli-agent[voice,deepgram]'
# or, in a checkout:
uv sync --extra voice --extra deepgram

You’ll also need:

• libopus on the system — apt install libopus0 (Debian/Ubuntu), brew install opus (macOS), bundled on Windows via discord.py’s wheel.
• ffmpeg on PATH if you want to use voice play (file/URL playback) or TTS playback — apt install ffmpeg / brew install ffmpeg / winget install ffmpeg.

For provider-specific extras:

Step 2 — Run the doctor

Before you debug “nothing is happening”, run:

discli doctor

doctor reports the state of every piece the voice stack needs. A working voice install looks like:

CORE
  [ok] python — Python 3.12.1
  [ok] bot token — configured
  [ok] discord.py — v2.7.1

VOICE
  [ok] libopus — loaded
  [ok] PyNaCl — v1.5.0
  [ok] discord-ext-voice-recv — v0.5.2a179
  [ok] davey (DAVE crypto) — v0.1.5
  [ok] DAVE/Opus patches — install ok

STT
  [ok] DEEPGRAM_API_KEY — set
  [--] OPENAI_API_KEY (Whisper) — not set

TTS
  [--] ELEVENLABS_API_KEY — not set
  [--] OPENAI_API_KEY (TTS) — not set
  [ok] DEEPGRAM_API_KEY (Aura) — set

TOOLS
  [ok] ffmpeg — /usr/bin/ffmpeg

OK — no problems found (3 optional check(s) skipped).

Use --json for machine-readable output (handy in CI):

discli doctor --json

Step 3 — Join a voice channel

Pick a voice channel ID (right-click → Copy ID in Discord with Developer Mode on, or use discli channel list), then:

discli voice join "<channel name or id>"

voice join is a one-shot command — it connects, confirms, and disconnects. To stay connected you need either an interactive session (next step) or discli serve (covered in Serve Mode).

Step 4 — Listen and transcribe

Start a transcription session in two terminals.

Terminal 1 — keep the bot in the channel:

discli voice join "general"

Terminal 2 — listen with Deepgram streaming STT:

export DEEPGRAM_API_KEY=...
discli voice listen

You’ll see lines like:

[123456789012345678] hello can you hear me
[987654321098765432] yes loud and clear

Each line is [user_id] text. Hit Ctrl+C to stop.

Step 5 — Speak with TTS

export ELEVENLABS_API_KEY=...
discli voice speak "joining the call now"

Pick a provider with --server config or env var:

Customise the voice ID with --voice <id> and the speech rate with --speed 0.8.

Step 6 — Play audio

discli voice play /path/to/file.mp3
discli voice play https://example.com/stream.opus

Playback goes through ffmpeg so any format ffmpeg understands works. Use voice stop, voice pause, voice resume to control it.

Step 7 — Build the meeting transcriber

discli ships a complete meeting-transcription example at examples/meeting_transcriber.py. It joins a server voice channel, transcribes everyone with their display names, appends to ~/.discli/transcripts/meeting-<timestamp>.md, and uses the Claude Agent SDK to produce a structured summary on Ctrl+C.

Set up:

pip install 'discord-cli-agent[voice,deepgram]' claude-agent-sdk
discli config set token YOUR_BOT_TOKEN
export DEEPGRAM_API_KEY=...

Run:

python examples/meeting_transcriber.py <voice_channel_id>

Output during the meeting:

Connected as MyBot#1234 (12345)
Listening to #standup. Transcript: /home/me/.discli/transcripts/meeting-20260514-103000.md
Press Ctrl+C to stop and generate a summary.

- **[10:30:14] Roy:** ok let's go around — what did everyone do yesterday
- **[10:30:22] Sara:** finished the auth migration, started on the rate limiter
- **[10:30:35] Roy:** nice

On Ctrl+C:

Stopping listener…
Generating summary from 47 line(s)…
Summary cost: $0.0084

=== Meeting Summary ===

## Summary
The team did a standup covering yesterday's work and today's plan.
Sara finished the auth migration; rate limiter is next.

## Key decisions
- Move forward with read-path-first for the migration this sprint.

## Action items
- Sara: finish rate limiter today.
- Roy: write the migration runbook by Friday.

## Open questions
- Do we backfill old sessions or expire them?

The summary is appended to the same transcript file. You can re-summarise later by piping the markdown back to Claude.

Behind the scenes — DAVE encryption

Modern Discord wraps voice in DAVE (Discord Audio and Video Encryption — their end-to-end voice encryption). The off-the-shelf discord-ext-voice-recv library strips the legacy SecretBox layer but doesn’t know about DAVE, so libopus rejects every packet as corrupted stream.

discli patches PacketDecoder._decode_packet at runtime to insert a davey.DaveSession.decrypt(...) call between the SecretBox layer and libopus, and wraps pop_data so a single bad packet can’t kill the listener. The patches install lazily on the first VoiceEngine.listen_start(...) call — you don’t need to do anything.

If discli doctor reports [FAIL] DAVE/Opus patches, voice listening is broken at the framework level. File an issue with the doctor output.

Permissions

The chat profile denies voice — agents using chat cannot join or speak. Use voice, moderation, or full to allow voice actions:

discli --profile voice voice join general
# or
DISCLI_PROFILE=voice discli voice listen

The readonly profile allows the stateless voice lookups (voice status, voice where, voice members) so a read-only agent can answer “who’s in voice?” without being able to join.

See Permission Profiles for full details.

Troubleshooting

See the full list in Common Issues.

What’s next

• Read the Meeting Transcription use case for a deeper walkthrough of the example.
• Build a persistent agent that does voice + text — see Building Agents and Serve Mode.
• Reference all voice commands in CLI Commands.
• See voice serve actions and events in Serve Actions and Serve Events.