Repolode
Submit
ysharma3501

ysharma3501/LuxTTS

Lightweight zipvoice-based TTS model for voice cloning and 48kHz speech generation, running at over 150x realtime on GPU.

Open-source alternative to:ElevenLabs

View on GitHub
Stars
5,487
Forks
685
Open issues
32
Last commit
5 months ago

What is it?

LuxTTS is a text-to-speech model built on the ZipVoice architecture and distilled to 4 sampling steps. Given a short reference recording, it clones the speaker's voice and generates speech from text. Output is 48kHz, using a custom vocoder where the default ZipVoice setup uses 24kHz.

The project targets speed and low memory use. It is described as running at up to 150x realtime on a single GPU, faster than realtime on CPU, and fitting within about 1GB of VRAM. Supported devices include CUDA, CPU, and MPS for Macs.

Usage is a Python API: load the model with LuxTTS(...), encode a prompt audio file with encode_prompt, then call generate_speech with optional parameters such as num_steps, speed, and t_shift. The repository also links a Hugging Face model and Space, a Colab notebook, and community integrations including a Gradio app, ComfyUI nodes, and an ONNX port.

Key features

  • 01Voice cloning from a short reference audio file (wav or mp3, at least about 3 seconds).
  • 0248kHz speech output.
  • 03Up to 150x realtime on a single GPU and faster than realtime on CPU.
  • 04Runs within about 1GB of VRAM.
  • 05Supports CUDA, CPU, and Apple MPS devices.

Use cases

  • Generating voiceover or narration in a cloned voice on a local GPU.
  • Running voice synthesis on a laptop CPU or a Mac.
  • Building a TTS backend for apps that need fast, high-sample-rate speech.

Quick start

  1. 1

    Clone the repository

    git clone https://github.com/ysharma3501/LuxTTS.git
  2. 2

    Enter the project folder

    cd LuxTTS
  3. 3

    Install dependencies

    pip install -r requirements.txt
  4. 4

    Load the model on GPU in Python and generate speech (see README inference examples)

    from zipvoice.luxvoice import LuxTTS

Things to know

  • Voice cloning needs a reference audio file of at least 3 seconds.
  • Speed claims of 150x realtime refer to a single GPU; the README gives no CPU or Mac benchmarks beyond saying CPU is faster than realtime.
  • Float16 inference is not yet released, so it currently runs in float32.

Topics

#text-to-speech#voice-cloning#speech-synthesis#pytorch#audio#deep-learning

FAQ

What is LuxTTS?
Lightweight zipvoice-based TTS model for voice cloning and 48kHz speech generation, running at over 150x realtime on GPU. LuxTTS is a text-to-speech model built on the ZipVoice architecture and distilled to 4 sampling steps. Given a short reference recording, it clones the speaker's voice and generates speech from text. Output is 48kHz, using a custom vocoder where the default ZipVoice setup uses 24kHz.
Is LuxTTS free to use? What license does it have?
LuxTTS is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is LuxTTS still maintained?
Updated within 6 months. The last commit to LuxTTS was on Jun 5, 2026.
What can I use LuxTTS for?
Generating voiceover or narration in a cloned voice on a local GPU. Running voice synthesis on a laptop CPU or a Mac. Building a TTS backend for apps that need fast, high-sample-rate speech.
How do I install LuxTTS?
Clone the repository: git clone https://github.com/ysharma3501/LuxTTS.git Enter the project folder: cd LuxTTS Install dependencies: pip install -r requirements.txt Load the model on GPU in Python and generate speech (see README inference examples): from zipvoice.luxvoice import LuxTTS
Where can I get LuxTTS?
The source code and installation instructions are on GitHub: https://github.com/ysharma3501/LuxTTS

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects