Install KorvaTTS with Conda

Introduction

KorvaTTS is an open-source text-to-speech system focused on Vietnamese with natural code-switching (Vietnamese sentences that contain English words, brand names, tech terms). It runs entirely on-device through ONNX Runtime: no GPU required, no API calls, 44.1 kHz output.

GitHub: https://github.com/dogenthq/KorvaTTS
Hugging Face: https://hf.co/collections/dogenthq/KorvaTTS
Demo Page: https://hf.co/spaces/hugging-apps/korvatts-demo
Try it online: https://heykorva.com

The model is our own re-implementation of the 99M-parameter Supertonic 3 architecture (flow-matching TTS), trained from scratch on Vietnamese data, with a vocoder based on BlueCodec. Upstream Supertonic is being archived by its authors; KorvaTTS continues the line with a Vietnamese focus and an open roadmap toward voice cloning and training.

🌐 Try it in your browser — no install: heykorva.com

Highlights

• 🇻🇳 Vietnamese-first — trained on Vietnamese audiobook speech plus a private code-switching set
• 🔀 Code-switching — English words inside Vietnamese sentences are read naturally
• ⚡ Fast on CPU — real-time factor well below 1 on a laptop CPU; GPU optional
• 🔊 44.1 kHz WAV output, no external upsampler
• 🧩 Drop-in compatible with Supertonic 3 runtimes (see Compatibility)
• 🪪 Apache-2.0 for both code and weights

Prerequisites

System requirements:

• Operating System: Windows 10/11 (64-bit), macOS, or Linux (Debian/Ubuntu).
• Python: version >= 3.10 required
• Disk Space: 5GB+ 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.

ONNX Runtime doesn’t publish a separate build per cuXXX tag the way PyTorch does. There are only three CUDA builds, so your list collapses into three groups:

Video tutorial

Coming soon!

Step 1. Install Miniconda Package

Step 2. Create Conda Environment

Create an environment.yml file

name: korvatts
channels:
- conda-forge
- defaults

dependencies:
# Python version (requires >= 3.10)
- python=3.10

- pip
- pip:
    # Install KorvaTTS directly from PyPI
    - korvatts

name: korvatts
channels:
- conda-forge
- defaults

dependencies:
# Python version (requires >= 3.10)
- python=3.10

- pip
- pip:
    # GPU version of ONNX runtime
    # This is for < CUDA 12.x
    # - onnxruntime-gpu==1.20.*

    # This is for CUDA 12.x
    - onnxruntime-gpu[cuda,cudnn]<1.27

    # This is for CUDA 13.x
    # - onnxruntime-gpu[cuda,cudnn]>=1.27

    # Install KorvaTTS directly from PyPI
    - korvatts

name: korvatts
channels:
- conda-forge
- defaults

dependencies:
# Python version (requires >= 3.10)
- python=3.10

- pip
- pip:
    # Install KorvaTTS directly from PyPI
    - korvatts

Activate conda environment:

conda env create -f environment.yml
conda activate korvatts

Download weights

korvatts voices

Step 3. Run the Inference

Command Line Interface

Now, run the basic example command to synthesize speech:

korvatts synth "Xin chào, đây là KorvaTTS." -v khanh_vy -o hello.wav

The result will be the audio file output.wav

Python Script

Create a file named korva_test.py:

from korvatts import TTS

tts = TTS()  # first run downloads weights + voices from Hugging Face
print(tts.list_voices())

wav, duration = tts.synthesize(
    "Hôm nay team mình sẽ demo tính năng text-to-speech mới.",
    voice="khanh_vy",
    lang="vi",
    total_steps=32,  # default; best quality. Lower (8-16) for faster synthesis
    speed=1.05,      # default; 0.5 (slow) .. 2.0 (fast)
)
tts.save_audio(wav, "output.wav")

Run the script to convert text to speech:

python korva_test.py

The result will be the audio file output.wav

Error and Warning

Notes

• Correction to my earlier advice: the newer Hugging Face login command is hf auth login, not huggingface-cli login.