AI Call Summarizer Discord Bot

A Discord bot that joins a voice channel, records each speaker separately, transcribes the audio with speaker labels, and produces a Markdown summary with TL;DR, key points, and action items — all via a configurable LLM (Ollama by default) and a configurable speech-to-text provider (local whisper.cpp by default, OpenAI Whisper API as alternative).

How it works

Discord exposes each voice-channel participant as a separate Opus stream, so speaker labels come for free — no diarization model needed. Per-user audio is captured to PCM, transcribed independently, then stitched chronologically into a single transcript that is fed to the LLM.

Prerequisites

• Node.js 18.17+
• A Discord application + bot token (https://discord.com/developers/applications)
  • Required bot intents: Server Members, Message Content is NOT needed.
  • Required permissions: Connect, Speak, Send Messages, Attach Files, Use Slash Commands.
• Ollama running somewhere reachable (default http://localhost:11434) with a model pulled (e.g. ollama pull llama3.1).
• For local STT: whisper.cpp built, with the whisper-cli (or main) binary on PATH plus a GGML model file (e.g. ggml-base.en.bin).
• On Windows, @discordjs/opus and better-sqlite3 are native modules — you will need Visual Studio Build Tools (C++) installed for npm install to compile them.

Install

npm install
cp .env.example .env
# fill in DISCORD_TOKEN and DISCORD_APP_ID
npm start

On first run, the bot registers its slash commands (to DISCORD_DEV_GUILD_ID if set, otherwise globally) and creates data/bot.db.

Step-by-step Windows setup

1. Ollama (local LLM)

1. Download and install Ollama from https://ollama.com/download.
2. Pull a model (fast, ~4.7 GB for llama3.1 8B):

   ollama pull llama3.1

3. Verify it's running:

   curl http://localhost:11434/api/tags

   Should return JSON listing your models.

2. whisper.cpp (local STT)

1. Open https://github.com/ggerganov/whisper.cpp/releases and download the latest Windows release zip (look for a file like whisper-bin-x64.zip — pick the CUDA variant if you have an NVIDIA GPU).

2. Extract to a path with no spaces, e.g. C:\tools\whisper.cpp\. You should see whisper-cli.exe (or main.exe in older releases) inside.

3. Add the folder to PATH or set WHISPER_CPP_BIN in your .env to the full path of the binary, e.g.:

   WHISPER_CPP_BIN=C:\tools\whisper.cpp\whisper-cli.exe
   

4. Download a GGML model from https://huggingface.co/ggerganov/whisper.cpp/tree/main. Good starter choices:

   Model	Size	Notes
   ggml-base.en.bin	142 MB	English-only, fast, decent accuracy
   ggml-small.en.bin	466 MB	English-only, better accuracy
   ggml-medium.en.bin	1.5 GB	Excellent English, slower
   ggml-large-v3.bin	3.1 GB	Multilingual, best accuracy, needs a GPU

   Save it next to the binary, e.g. C:\tools\whisper.cpp\ggml-base.en.bin.

5. Quick sanity check:

   whisper-cli -m C:\tools\whisper.cpp\ggml-base.en.bin -f samples\jfk.wav -oj

   Should print segments and write a .json next to the input.

3. Configure the bot in Discord

Run these in your server (Manage Server permission required):

/config channel channel:#summaries
/config stt provider:whispercpp model_path:C:\tools\whisper.cpp\ggml-base.en.bin
/config llm provider:ollama base_url:http://localhost:11434 model:llama3.1
/help

/help will show a live check of what's configured and what's missing.

4. Record a call

/join            (while you're in a voice channel)
... talk ...
/leave

The summary + full transcript.txt will be posted to your configured summary channel.

Slash commands

Recording

Command	What it does
/join	Bot joins your current voice channel and starts recording every speaker.
/leave	Stops recording, transcribes, summarizes, posts to the configured channel.
/help	Shows setup instructions and a live config-status check.

Configuration

Command	What it does
/config llm provider:<ollama> base_url:<url> model:<name>	Sets the LLM provider, endpoint, and model.
/config stt provider:<whispercpp|openai> model_path:<path> api_key:<key>	Sets the transcription provider and options.
/config channel channel:<#channel>	Sets the channel where summaries are posted.
/config role role:<@role>	Allows a role to run /config (in addition to Manage Server admins).
/config show	Shows current configuration (API key masked).

Anyone with Manage Server — or the configured role — may use /config.

Defaults

Setting	Default
LLM provider	ollama
LLM base URL	http://localhost:11434
LLM model	llama3.1
STT provider	whispercpp
STT model path	(unset — must configure)
Summary channel	(unset — must configure)

Security notes

• The OpenAI API key is stored plaintext in SQLite. If you run this bot on a shared host, restrict filesystem access to data/bot.db.
• /config show masks the API key in its response, but the DB file itself is unencrypted.
• Anyone with Manage Server (or the configured role) can change provider URLs and models — verify your role setup before running in large servers.

Out of scope (v1)

• Real-time transcription while the call is ongoing.
• LLM providers other than Ollama (the adapter layer makes it straightforward to add OpenAI/Anthropic/etc. — see src/llm/).
• Encryption of stored credentials at rest.
• Per-guild concurrency limits.