Install OmniVoice TTS on Windows with Conda

Introduction

OmniVoice is a state-of-the-art massively multilingual zero-shot text-to-speech (TTS) model supporting over 600 languages. Built on a novel diffusion language model-style architecture, it generates high-quality speech with superior inference speed, supporting voice cloning and voice design.

GitHub: https://github.com/k2-fsa/OmniVoice
Hugging Face: https://hf.co/k2-fsa/OmniVoice
Paper: https://arxiv.org/abs/2604.00688
Demo: https://hf.co/spaces/k2-fsa/OmniVoice
Audio Samples: Coming soon!

Prerequisites

System requirements:

• Operating System: Windows 10/11 (64-bit), macOS, or Linux (Debian/Ubuntu).
• Python: version >= 3.10 required
• Disk Space: 10GB+ recommended (for dependencies and model cache). At least 400 MB for Miniconda; 3 GB+ for full Anaconda.
• The GPU is optional but HIGHLY Recommended for Performance
• Internet: For downloading dependencies and models from Hugging Face Hub.

Note: CUDA version check by command

nvidia-smi

Video tutorial

Coming soon!

Step 1. Install Miniconda Package

Download Miniconda: https://www.anaconda.com/download/success?reg=skipped

Direct link: https://anaconda.com/api/installers/Miniconda3-latest-Windows-x86_64.exe

Step 2. Create Conda Environment

Create a conda environment:

name: omnivoice
channels:
  - pytorch
  - conda-forge
  - defaults
dependencies:
  # Python version >= 3.10 required
  - python=3.10
  - pip
  - pip:
      # PyTorch CUDA 12.6 wheels
      - --extra-index-url https://download.pytorch.org/whl/cu126
      - torch
      - torchaudio

      # OmniVoice-TTS
      - omnivoice

Activate conda environment:

conda env create -f environment.yml
conda activate omnivoice

Step 3. Create a Python Script

Create a file named omnivoice_tts_test.py:

from omnivoice import OmniVoice
import soundfile as sf
import torch

model = OmniVoice.from_pretrained(
    "k2-fsa/OmniVoice",
    device_map="cuda:0",
    dtype=torch.float16
)
# Apple Silicon users: use device_map="mps" instead
# Intel Arc GPU users: use device_map="xpu" instead

audio = model.generate(
    text="Hello, this is a test of zero-shot voice cloning.",
    # ref_audio="ref.wav",
    # ref_text="Transcription of the reference audio.",
    instruct="female, young adult, high pitch, british accent",
) # audio is a list of `np.ndarray` with shape (T,) at 24 kHz.

# If you don't want to input `ref_text` manually, you can directly omit the `ref_text`.
# The model will use Whisper ASR to auto-transcribe it.

sf.write("out.wav", audio[0], 24000)

Run it:

python omnivoice_tts_test.py

The result will be the audio file out.wav

Launch the local web UI

Try OmniVoice without coding:

omnivoice-demo --ip 0.0.0.0 --port 8001

Open your browser and navigate to http://127.0.0.1:8001. The system will automatically download the required model weights from HuggingFace during this first run.