---

Celune is a real-time AI TTS character engine focused on expressive voice delivery with a well-made user experience.

It has been designed to provide fast, high-quality voice generation, even on consumer-grade hardware.

It was proudly made in 🇵🇱 for your listening pleasure.

Features

• Real-time buffered speech generation pipeline
• Distinct default voice styles: calm, balanced, bold, upbeat
• Multiple operation modes: frontend, API and extensions
• Stable long-form narration with low risk of drifting
• Native audio controls & effects via built-in DSP
• Optimized GPU inference where possible
• Configurable character voices via CEVOICE voice packs
• Characters can respond back to you

Note on development

Celune is against the stance of "vibe coding" used in development.

None of the 15,000+ lines of code in Celune were created solely using AI. AI tools (e.g. Codex) were only used to assist in faster development, iteration and solving issues.

All decisions and implementations were reviewed, validated, and approved by human developers.

Celune never was, and will never become an "AI slop" project.

Voices & samples

Each voice is demonstrated using a short introduction and a longer narration sample to showcase consistency, pacing, and expressiveness.

Qwen

Voice	Intro	Narration
Balanced	▶️ Play	▶️ Play
Calm	▶️ Play	▶️ Play
Bold	▶️ Play	▶️ Play
Upbeat	▶️ Play	▶️ Play

VoxCPM2

Voice	Intro	Narration
Balanced	▶️ Play	▶️ Play
Calm	▶️ Play	▶️ Play
Bold	▶️ Play	▶️ Play
Upbeat	▶️ Play	▶️ Play

Celune Mini

Voice	Intro	Narration
Balanced	▶️ Play	▶️ Play
Calm	▶️ Play	▶️ Play
Bold	▶️ Play	▶️ Play
Upbeat	▶️ Play	▶️ Play

The demonstration lines try to showcase Celune's best, but they may include minor mistakes. This is an inherent limitation with TTS models, and Celune should not be blamed for it.

These examples were recorded with the default voice pack, which contains canonical Celune voices. Other voice packs may be created and installed by the end user, allowing for other characters to speak through Celune.

Browse the demos directory for demonstration content from the current version of Celune, as well as any past releases.

Caution

Do not use markup or unknown tags (e.g. <...>). They may be interpreted as control sequences and break speech output. Refer to the model's known tags before including them.

Do not mix multiple languages in one sentence.
Keep language boundaries clear and explicit.

Good:

This is a sentence. This is another sentence. [laughter]

Bad:

<think>Thinking text.</think>
This is a sentence, 中文, 日本語, 한국어.

Samples were captured directly from Celune's output directory. No extra post-processing was applied.

For details on voice production, check the VOICES.md file.

System Requirements

Celune requires Python 3.12 or 3.13.

Celune also depends on external system dependencies that are not available in pip:

• CUDA Toolkit 12.8 - only if not using pre-built PyTorch wheels
• SoX (Sound eXchange) - required for audio processing
• Rubber Band library - required to control Celune's speed
• OpenRGB - required to glow compatible devices
• Symbolic link support - recommended on Windows for optimal operation
• C/C++ compiler - to compile required dependencies for VoxCPM2

Celune requires an RTX 30 series GPU or newer to use most features.

CPU-only execution is only supported with Celune Mini.

Usage of Celune's UI requires an ANSI-capable terminal. Non-compliant terminals can only use the headless (CEF) mode.

The terminal should support True Color, especially when using voice packs that declare new app themes.

Terminals not supporting True Color may look incorrect, as Textual will fall back to a lower color mode.

If Rubber Band is not installed or fails to run, Celune will speak at normal speed, and speed controls will be unavailable.

VRAM presets & requirements

Celune has several VRAM presets available. Here are their baseline requirements:

Low (6GB VRAM recommended)

This preset supports:

• Qwen3-TTS 0.6B / Celune Mini
• Normalization on CPU

Medium (8GB VRAM required)

This preset supports:

• All of the above, and:
• Qwen3-TTS 1.7B

High (12GB VRAM required)

This preset supports:

• All of the above, and:
• VoxCPM2
• Persona (4-bit)

Extra High (16GB+ VRAM recommended)

This preset supports:

• All of the above, and:
• Persona (8-bit)
• Normalization on GPU
• Upcoming advanced features

The desired preset may be set in Celune's configuration file. Refer to default_config.yaml for details.

---

Performance may be reduced when running GPU intensive applications along with Celune.

Tested on: RTX 5070 (12 GB VRAM)

Installation

# Download Celune
git clone https://github.com/celunah/celune
cd celune

# Quick setup
python setup.py

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

# Or on Unix systems:
curl -Ls https://astral.sh/uv/install.sh | sh

# Validate uv works
uv --version

# Expected output:
# uv 0.11.2 (02036a8ba 2026-03-26 x86_64-pc-windows-msvc) (or similar version)

# Create environment
# Celune expects Python 3.12 or 3.13
uv sync

# Run
# Command Prompt users
celune

# PowerShell users
.\celune.exe

# Or on Unix systems:
./celune.AppImage

You can also open Celune from within your desktop by running the aforementioned executables. They are usable as an entry point.

SoX & Rubber Band installation

If SoX & Rubber Band are already installed, you can skip this section.

Windows (Scoop)

# Install Scoop if you don't already have it
powershell -ExecutionPolicy RemoteSigned -c "irm https://get.scoop.sh | iex"

# Install SoX
scoop install sox

# Install Rubber Band
scoop install rubberband

Linux (Debian/Ubuntu)

sudo apt install sox rubberband-cli

Linux (Arch Linux)

sudo pacman -S sox rubberband

Validate SoX & Rubber Band are installed

sox --version

# Expected output:
# sox:      SoX v14.4.2 (or similar version)

rubberband --version

# Expected output:
# 4.0.0 (or similar version)

OpenRGB installation

To install OpenRGB, go to https://openrgb.org/, download and install a package appropriate for your platform. This will allow Celune to glow up your PC as she speaks.

C/C++ compiler setup

Celune's VoxCPM2 backend may require a C/C++ compiler to compile dependencies. To install a suitable compiler, run one of the following commands:

This is not required to use the Qwen backend, but you may need to install dependencies manually.

# Windows
winget install Microsoft.VisualStudio.2022.BuildTools --override "--wait --passive --add Microsoft.VisualStudio.Workload.VCTools --includeRecommended"

# Linux (Ubuntu)
sudo apt install build-essential

# Linux (Arch Linux)
sudo pacman -S base-devel

CUDA Toolkit 12.8 installation

This step can be skipped if you are using pre-built PyTorch wheels.

Download and install CUDA Toolkit 12.8 from NVIDIA:

https://developer.nvidia.com/cuda-12-8-0-download-archive

Make sure to:

• Select the correct OS and version
• Install both CUDA Toolkit and NVIDIA drivers (if not already installed)

Make sure you install version 12.8, as Celune does not work with older or newer versions of the toolkit.

After installation, verify CUDA:

nvidia-smi

You should see your GPU listed along with driver information.

Symbolic links (Windows)

Symbolic links are recommended for best performance and compatibility.

To enable them:

• Enable Developer Mode in Windows settings
  (Settings → Privacy & Security → For Developers)

Without this, Celune may require elevated permissions or fall back to slower behavior.

REST API

See API.md for REST API configuration, authentication, endpoints, and cURL examples. The API allows programmatic usage of all Celune features. It can be used both as a public and local interface.

Screenshots

The below images showcase Celune's user interface.

Before init

Ready

Talking

Change voice

Commands

Extension invoke

"Your voice, your way."