Yap To Text

Real-time captions for VRChat and Twitch, powered by Qwen3-ASR.

Turn your voice into text displayed inside VRChat's chatbox, or send transcriptions to a Twitch channel — no typing required.

Features

• Live VRChat captions — sends real-time transcriptions directly to VRC's in-game chatbox via OSC
• Twitch mode — stream your own live captions to Twitch chat instead (--twitch)
• OBS browser source — built-in web server serves live subtitles for OBS Studio overlays (run with --twitch)
• Clipboard copy — dump full transcriptions to your clipboard (--clipboard)
• Hallucination & erase detection — filters ASR hallucinations and clears the buffer on an erase keyword

System Requirements

Requirement	Details
OS	Linux or Windows
GPU	NVIDIA with 8+ GB VRAM (CUDA 12.8+)
Python	3.12
Package manager	uv

Installation

# Clone the repo
git clone https://github.com/atomicflag/yap-to-text.git
cd yap-to-text

# Create a virtual environment and install dependencies
uv sync

That's it — uv resolves everything including the PyTorch and flash-attn wheels. The Qwen3-ASR model (1.7B) downloads automatically on first run.

Note: If you're not familiar with Python or the uv ecosystem, see uv getting started for a quick intro.

Configuration (optional)

Copy the example config and fill in your values:

cp config.json.example config.json

VRChat (OSC) mode — no extra setup needed

Default mode sends text to VRChat's chatbox over localhost OSC. Just make sure VRChat is running.

Twitch mode — requires a Twitch Developer app

1. Register an app on the Twitch Developer Console
2. Note your Client ID and Client Secret
3. Add them to config.json under twitch, along with your channel name
4. On first run with --twitch, a browser window opens for OAuth authorization — grant permissions and you're done

{
  "twitch": {
    "app_id": "your_client_id",
    "app_secret": "your_client_secret",
    "target_channel": "your_channel_name"
  },
  "hallucinations": ["The.", "."],
  "erase_keyword": "not what i said",
  "audio_input_device": null,
  "audio_output_device": null
}

Usage

VRChat mode (default)

uv run python -m ytt

Transcriptions appear in VRC's chatbox in real-time.

Twitch mode

uv run python -m ytt --twitch

Sends final transcriptions to Twitch chat and serves live captions via a local web server (port 9098) for OBS browser source. A blip sound plays when each speech segment ends.

To add the OBS browser source:

1. Open OBS → Sources → Browser
2. Set URL to http://localhost:9098
3. Adjust width/height as needed

Clipboard mode (works with either VRChat or Twitch)

uv run python -m ytt --clipboard    # or --twitch --clipboard

Copies the full running transcription to your system clipboard whenever a segment is finalized.

Options

Run uv run python -m ytt --help to see all available command-line flags:

Flag	Description
--twitch	Enable Twitch chat mode (sends to Twitch instead of VRC)
--clipboard	Copy final transcriptions to the system clipboard
--config PATH	Custom path to config.json (default: config.json)

License

MIT — see LICENSE for details.