cornfab

Text-to-speech Shiny app for the Cornball AI ecosystem.

Features

• Multi-backend TTS: Chatterbox, Qwen3-TTS, OpenAI, ElevenLabs, fal.ai
• Voice selection: 9 built-in Qwen3 voices, OpenAI voices, ElevenLabs library
• Voice cloning: Upload reference audio for Chatterbox/Qwen3 backends
• Voice design: Create custom voices from natural language descriptions (Qwen3)
• Save as voice: Save generated audio as a reusable voice for cloning
• History: Persistent storage with audio playback in ~/.cornfab/

Installation

# Install from GitHub
remotes::install_github("cornball-ai/cornfab")

Usage

library(cornfab)
run_app()  # Runs on port 7803

Backends

Backend	Type	Port	Features
Chatterbox	Container	7810	Voice cloning, exaggeration control
Qwen3-TTS	Container	7811	9 voices, voice design, 10 languages
OpenAI	API	-	6 voices, tts-1/tts-1-hd models
ElevenLabs	API	-	Large voice library, multilingual
fal.ai	API	-	F5-TTS, Dia, Orpheus models

Container Setup

Chatterbox and Qwen3-TTS run as local Docker containers. You must:

1. Download models before running containers
2. Start containers manually - cornfab does not auto-start containers

Model Storage

Models are stored in the HuggingFace cache:

~/.cache/huggingface/hub/

Mount this directory when running containers:

-v ~/.cache/huggingface:/root/.cache/huggingface

Downloading Models

Chatterbox

Model: ResembleAI/chatterbox (~2GB)

Option 1: R with hfhub (recommended)

# install.packages("hfhub")
hfhub::hub_snapshot("ResembleAI/chatterbox")

Option 2: Python with huggingface_hub

pip install huggingface_hub
python -c "from huggingface_hub import snapshot_download; snapshot_download('ResembleAI/chatterbox')"

Option 3: curl

mkdir -p ~/.cache/huggingface/hub/models--ResembleAI--chatterbox/snapshots/main
cd ~/.cache/huggingface/hub/models--ResembleAI--chatterbox/snapshots/main
curl -LO https://huggingface.co/ResembleAI/chatterbox/resolve/main/chatterbox.safetensors
curl -LO https://huggingface.co/ResembleAI/chatterbox/resolve/main/s3gen.safetensors
curl -LO https://huggingface.co/ResembleAI/chatterbox/resolve/main/t3_cfg.safetensors
curl -LO https://huggingface.co/ResembleAI/chatterbox/resolve/main/ve.safetensors

Qwen3-TTS

Three models are needed for full functionality:

• Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice - Built-in speakers (~7GB)
• Qwen/Qwen3-TTS-12Hz-1.7B-Base - Voice cloning (~7GB)
• Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign - Voice design (~7GB)

Option 1: R with hfhub (recommended)

# Download all three models
hfhub::hub_snapshot("Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice")
hfhub::hub_snapshot("Qwen/Qwen3-TTS-12Hz-1.7B-Base")
hfhub::hub_snapshot("Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign")

Option 2: Python with huggingface_hub

python -c "from huggingface_hub import snapshot_download; \
  snapshot_download('Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice'); \
  snapshot_download('Qwen/Qwen3-TTS-12Hz-1.7B-Base'); \
  snapshot_download('Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign')"

Option 3: Git LFS

cd ~/.cache/huggingface/hub
git lfs install
git clone https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice

Running Containers

Note: Containers default to LOCAL_FILES_ONLY=true and will fail if models aren't pre-downloaded. Set LOCAL_FILES_ONLY=false to enable auto-download (not recommended for production).

Chatterbox (port 7810)

# Build (if not using ghcr.io)
cd ~/chatterbox-tts-api
docker build -f docker/Dockerfile -t chatterbox-tts-api .

# Run
docker run -d --gpus all --network=host --name chatterbox \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -e PORT=7810 \
  chatterbox-tts-api

Qwen3-TTS (port 7811)

# Build (if not using ghcr.io)
cd ~/qwen3-tts-api
docker build -f Dockerfile.blackwell -t qwen3-tts-api:blackwell .

# Run (Blackwell GPUs - RTX 50xx)
docker run -d --gpus all --network=host --name qwen3-tts-api \
  -v ~/.cache/huggingface:/cache \
  -e PORT=7811 \
  -e USE_FLASH_ATTENTION=false \
  qwen3-tts-api:blackwell

# Run (older GPUs - Ampere, Ada Lovelace)
docker build -t qwen3-tts-api .  # Use default Dockerfile
docker run -d --gpus all --network=host --name qwen3-tts-api \
  -v ~/.cache/huggingface:/cache \
  -e PORT=7811 \
  qwen3-tts-api

API Backends

For OpenAI, ElevenLabs, and fal.ai, set environment variables:

export OPENAI_API_KEY="sk-..."
export ELEVENLABS_API_KEY="..."
export FAL_KEY="..."

Or configure in the app's API Settings panel.

Voice Design (Qwen3-TTS)

Qwen3-TTS supports creating custom voices from natural language descriptions.

VRAM Usage

Mode	Models Loaded	VRAM
Built-in voices	Base	~4.6GB
Voice cloning	+ CustomVoice	+4.1GB
Voice design	+ VoiceDesign	+3.9GB

Models load on first use and stay in memory. Restart the container to unload.

Recommended Workflow

To avoid keeping the VoiceDesign model loaded:

1. Design once: Check "Design voice from description", enter a description like "A warm, friendly female voice with a slight British accent", generate
2. Save as voice: Click "Save as Voice", enter a name (e.g., "warm-female")
3. Restart container: docker restart qwen3-tts-api to free ~5GB VRAM
4. Clone from saved: Select "warm-female (custom)" from the voice dropdown for future generations

Saved voices are stored in ~/.cornfab/voices/ and work with both Qwen3 and Chatterbox backends.

Development

# Clone
git clone https://github.com/cornball-ai/cornfab
cd cornfab

# Build and run
r -e 'tinyrox::document(); tinypkgr::install()'
r -e 'library(cornfab); run_app()'

Related

• tts.api - R package for TTS APIs
• earshot - Speech-to-text Shiny app