D&D Scribe Bot

A locally-hosted Discord bot that joins your voice channel, transcribes everyone in real time using Whisper, and generates structured D&D session notes using a local LLM via Ollama. No audio or transcript data ever leaves your machine.

Built on top of Discord-Realtime-STT-Bot with D&D-specific summarization, session management commands, and crash-safe rolling transcript logging.

---

How It Works

Discord Voice → discord-ext-voice-recv → STTAudioSink
    → Resample (48kHz stereo → 16kHz mono)
    → IPC Queue → STT Subprocess
        → Silero VAD (speech detection)
        → faster-whisper (speech-to-text)
    → Result Queue → Bot Process
        → Rolling transcript file (written live)
        → !summary / !leave → Ollama (LLM summarization)
            → Structured session notes posted to Discord
            → Markdown file saved to disk

The bot runs two processes:

• Bot process — handles Discord connection, voice receive, commands, and result polling
• STT subprocess — runs Silero VAD and faster-whisper in isolation so GPU inference never blocks the Discord event loop

Audio flows from Discord through a per-user resampler, into a multiprocessing queue, through voice activity detection, and into Whisper. Transcribed lines are written to both an in-memory list and a rolling file on disk as they arrive. At the end of a session, Ollama summarizes the full transcript into structured D&D notes.

Warning/Notice

It should go without saying, but despite Discord having DAVE, using this bot sort of negates it. You are transcribing basically everything said in the channel the bot is currently in. You are responsible for safely storing, using, and discarding these transcriptions. Make sure you have the permission of everyone in the channel and that they know they are being recorded. Everyone has the right to privacy and you should respect that when attempting to use this bot.

---

Requirements

Hardware

• NVIDIA GPU with 12GB+ VRAM recommended
• Tested on RTX 5070 Ti (16GB) — Blackwell architecture requires special setup (see below)

Software

• Windows 10/11 or Linux
• Python 3.11+
• uv (package manager)
• Ollama (local LLM server)
• FFmpeg
• NVIDIA driver 572.16+ (for RTX 5000 series / Blackwell)

---

Setup

1. Install Prerequisites

Ollama — download and install from ollama.com, then pull a model:

ollama pull qwen3:14b

FFmpeg — install via winget on Windows:

winget install ffmpeg

uv — install the package manager:

powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

2. Clone and Configure

git clone https://github.com/yourusername/dnd-scribe.git
cd dnd-scribe
cp .env.example .env

Edit .env with your settings — at minimum, set your DISCORD_TOKEN.

3. Install Dependencies

uv sync

PyTorch — CUDA setup (important)

PyTorch must be installed from the PyTorch index, not PyPI, to get CUDA support. The right CUDA version depends on your GPU:

GPU Generation	CUDA Version	Command
RTX 5000 series (Blackwell)	cu130	See below
RTX 3000/4000 series	cu128	--index-url https://download.pytorch.org/whl/cu128

For RTX 5000 series (Blackwell):

uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130

Verify PyTorch can see your GPU:

uv run python -c "import torch; print(torch.cuda.is_available()); print(torch.cuda.get_device_name(0))"

You should see True and your GPU name.

4. Create a Discord Bot

1. Go to discord.com/developers/applications
2. Click New Application
3. Go to the Bot tab → click Reset Token → copy the token into your .env
4. Under Privileged Gateway Intents, enable all three:
   • Presence Intent
   • Server Members Intent
   • Message Content Intent
5. Go to OAuth2 → URL Generator → check bot scope → check these permissions:
   • Read Messages / View Channels
   • Send Messages
   • Connect
   • Speak
6. Copy the generated URL, open it in your browser, and invite the bot to your server

5. Run

uv run main.py

You should see:

INFO  │ STT process spawned (PID ...)
INFO  │ Silero VAD loaded successfully.
INFO  │ Faster-Whisper loaded on CUDA.
INFO  │ STT Processor ready
INFO  │ Logged in as D&D Scribe#1234

---

Configuration

All settings are in .env. Copy .env.example to get started.

Variable	Default	Description
DISCORD_TOKEN	—	Required. Your bot token from the Developer Portal
COMMAND_PREFIX	!	Prefix for bot commands
STT_MODEL_ID	deepdml/faster-whisper-large-v3-turbo-ct2	Whisper model to use
STT_DEVICE	cuda	cuda or cpu
STT_COMPUTE_TYPE	int8_float16	Use int8_float16 for Blackwell GPUs, float16 for others
STT_LANGUAGE	en	Language code for Whisper transcription
STT_BEAM_SIZE	1	Higher = more accurate but slower
STT_MIN_AUDIO_BYTES	8000	Minimum audio chunk size before transcription (~0.25s)
OLLAMA_URL	http://localhost:11434	Ollama server URL
OLLAMA_MODEL	llama3.1:8b	Model to use for summarization
OLLAMA_TIMEOUT	180	Seconds to wait for Ollama response
LOG_LEVEL	INFO	Logging verbosity

Choosing an Ollama Model

For summarization quality vs VRAM fit on a single GPU:

Model	VRAM (Q4)	Notes
llama3.1:8b	~5GB	Fast, decent quality
qwen3:14b	~9GB	Recommended — better instruction following
mistral-small:22b	~14GB	Strong quality, tight fit on 16GB alongside Whisper

For Qwen3 models, add /no_think at the top of the prompts in bot/summarizer.py to disable the chain-of-thought reasoning mode, which is unnecessary for structured note generation.

Customizing the D&D Prompt

Open bot/summarizer.py and edit the CAMPAIGN_CONTEXT variable at the top:

CAMPAIGN_CONTEXT = "The party: Aldric (human paladin), Zessa (tiefling warlock), Pip (halfling rogue). Campaign: Curse of Strahd, currently in Barovia."

The more context you add here, the better the notes — the model has no memory of previous sessions, so anything you bake into the prompt improves how well it identifies characters and plot threads.

---

Commands

Command	Description
!join	Join your current voice channel and start transcribing
!leave	Flush audio, dump raw transcript, post session notes, and disconnect
!summary	Generate and post session notes without stopping the recording
!dumptranscript	Save the raw transcript to a file immediately, no summarization
!clearnotes	Wipe the in-memory transcript and start fresh from this point
!status	Show how long the session has been running and how many lines are captured

---

Output Files

All files are saved to the sessions/ directory in your project folder.

File	When created	Contents
rolling_<timestamp>.txt	On !join, written continuously	Every transcription line as it arrives — crash-safe
transcript_<timestamp>.txt	On !leave, !summary, !dumptranscript	Full raw transcript dump
session_<timestamp>_final.md	On !leave	Structured session notes from Ollama
session_<timestamp>_checkpoint.md	On !summary	Mid-session notes snapshot

The rolling file is the most important for crash safety — it's written and flushed to disk after every transcription line, so a Ctrl+C or crash loses at most the line being written at that exact moment.

---

Known Issues and Fixes Applied

Discord DAVE end-to-end encryption Discord mandated E2EE for all voice channels in early 2026, breaking every Python voice receive library. This bot uses a specific fork of discord-ext-voice-recv (rdphillips7/discord-ext-voice-recv@ddd2860) that implements DAVE decryption, along with the davey package. This is pinned in pyproject.toml.

Blackwell GPU (RTX 5000 series) + cuBLAS Blackwell requires CUDA 13.0 and the cu130 PyTorch build. Even with the correct build, float16 compute type may fail to find cublas64_12.dll. Set STT_COMPUTE_TYPE=int8_float16 in your .env to avoid cuBLAS entirely while keeping GPU inference.

Whisper hallucinations Whisper sometimes hallucinates short phrases ("Thank you.", "Thanks for watching.") on silence or noise that passes VAD. A blocklist in stt/transcriber.py drops known hallucinations, and STT_MIN_AUDIO_BYTES=8000 raises the minimum chunk size to reduce false VAD triggers.

Ollama context window for long sessions Long D&D sessions (90+ minutes) can produce transcripts well over 15,000 tokens. If the transcript exceeds num_ctx, Ollama silently truncates the prompt — including the instructions — producing garbage output. The summarizer chunks long transcripts, summarizes each part, then combines them in a final pass. num_ctx is set to 16384 by default.

---

Project Structure

dnd-scribe/
├── main.py              # Entry point — spawns STT process, starts bot
├── config.py            # All configuration, loaded from .env
├── .env.example         # Copy to .env and fill in your settings
├── bot/
│   ├── client.py        # Discord bot, commands, result polling, session management
│   ├── audio_sink.py    # Receives Discord audio, resamples, forwards to STT queue
│   └── summarizer.py    # Ollama integration, chunked summarization for long sessions
├── audio/
│   ├── resampler.py     # 48kHz stereo → 16kHz mono converter (torchaudio)
│   └── ring_buffer.py   # Pre-speech context buffer for VAD
├── stt/
│   ├── processor.py     # STT subprocess main loop
│   ├── transcriber.py   # faster-whisper wrapper with hallucination filtering
│   ├── vad.py           # Silero VAD wrapper
│   └── user_state.py    # Per-user audio buffers and state
├── utils/
│   └── logging.py       # Shared logging configuration
└── sessions/            # Output directory (auto-created)

---

Platform Support

This project has been developed and tested exclusively on Windows 11. It should work on Linux in principle — the dependencies are all cross-platform and the code doesn't use anything Windows-specific — but it has not been tested and no Linux setup instructions are provided. Things that may need attention on Linux:

• The uv install command differs (curl-based rather than PowerShell)
• FFmpeg is typically installed via your package manager (apt, pacman, etc.) rather than winget
• PyTorch CUDA builds and library paths may behave differently
• The multiprocessing spawn method defaults differ between Windows and Linux, which could affect the STT subprocess

Linux PRs are welcome. If you get it running, please open an issue or PR with any fixes needed.

---

AI Assistance

This project was built with significant assistance from Claude (Anthropic). The initial architecture, all source code, debugging, and documentation were developed collaboratively through an iterative back-and-forth process — identifying the right libraries, working through the DAVE encryption problem, diagnosing GPU and dependency issues, and refining the summarization pipeline.

The core STT architecture was adapted from an existing open-source project (see Credits), but the D&D-specific summarization layer, session management commands, crash-safe transcript logging, and chunked multi-pass summarization for long sessions were designed with Claude and implemented by me.

---

Credits

Core STT architecture adapted from Leehyunbin0131/Discord-Realtime-STT-Bot.

DAVE decryption via rdphillips7/discord-ext-voice-recv, implementing the fix from imayhaveborkedit/discord-ext-voice-recv#54.