ZONOS2

ZONOS2

Turns text into natural speech in more than 30 languages, with voice cloning from a sample, using a mixture-of-experts model for low latency. Runs locally on a Linux machine with an NVIDIA GPU.

Made by Zyphra on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: ZONOS2 is Zyphra's speech synthesis model. Its training data runs past 6 million hours of recordings in many languages. It also clones a voice from a short sample. The authors say its quality matches or beats leading TTS providers, with low delay thanks to a mixture-of-experts design.

Who it suits: It suits developers and creators who want speech generated on their own machine, in many languages, and who have a Linux box with an NVIDIA GPU.

Industry | Who uses it | How they use it | Example
Media | Video and podcast producers | Generate voice-overs from scripts | A narrated explainer from a written script
Software | App developers | Add speech output through a local server or Python | A reader feature that speaks article text
Localization | Content teams | Produce speech in several languages | The same announcement in English and Japanese
Accessibility | Tool builders | Read text aloud with a chosen voice | A screen reader that uses a cloned voice with permission

You can skip it if you have no NVIDIA GPU or no Linux machine, since the card says the local setup is Linux only on x86_64 with CUDA.

Where to use it:

• A Linux x86_64 machine with an NVIDIA GPU and a matching CUDA toolkit
• The included TTS inference server and its web page
• A Python script that calls the model directly, with no server
• The vendor's hosted version, which the card links for trying it without setup

Use case | Who | What they get back
Generate speech from text | Creators | An audio file of the spoken text
Clone a voice | Producers | Speech in a voice taken from a sample you provide
Batch generation in Python | Developers | Several audio files from a list of prompts

How to set it up:

• Linux on x86_64 with an NVIDIA GPU and a CUDA toolkit matching your driver
• The uv package tool
• The ZONOS2 project code from its repository

1. Download the model with Hugging Face's own steps, or let the server fetch it.
2. Get the project code and sync its environment with uv.
3. Start the TTS server with the model path.

pip install -U huggingface_hub
hf download Zyphra/ZONOS2

uv sync

How to use it: You start the server, then open its web page on your own machine, type text and pick a voice. You can also call the model from Python for offline runs. The server can scan a folder of speaker recordings and offer them as cloning voices.

uv run python -m minisgl --model-path Zyphra/ZONOS2 --tts-default-voices-dir ./default_voices/

from zonos2.message import TTSSamplingParams
from zonos2.tts import TTSLLM
tts = TTSLLM(model_path="Zyphra/ZONOS2")
results = tts.generate(
    ["Hello from the offline Python API.", "Batched prompts work too."],
    TTSSamplingParams(seed=42),
)
for i, result in enumerate(results):
    print(f"frames={len(result['audio_tokens'])}, eos_frame={result['eos_frame']}")
    tts.save_audio(result["audio"], f"output_{i}.wav")

Example: A person starts the server and puts a short recording of their own voice in the voices folder. They open the web page, type a sentence and generate it. What comes back is that sentence spoken in a voice that resembles the sample. Use only voices you have the right to clone.

For newcomers and experts: If you are new to this, try the vendor's hosted version first to hear the quality, then set up the local server. Keep the first test to one short sentence in English.

If you know the field, the card says inference feeds normalized UTF-8 bytes and a speaker embedding into a mixture-of-experts backbone that produces audio codec tokens. The local server builds on Mini-SGLang. Seeds are set through the sampling parameters in the Python API.

Start the TTS server
Provide text and a voice sample
The model generates audio tokens
The codec turns them into sound
Save or play the audio

Good to know:

• Languages: the card ranks them in three tiers, with 3 in the top tier, 10 in the second and 21 in the third.
• Limits: the local setup is Linux only on x86_64 and needs an NVIDIA GPU with CUDA.
• Voices: voice cloning can copy a real person's voice, so use only voices you have the right to use.
• License: Apache 2.0.
• Cost to run: no price is named for the local setup, and the card also links a hosted version from the vendor.
• Data it reaches: it copies model files to your computer, and the voice samples and text you give it stay on the machine running the server.

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: ZONOS2
• Downloads in the last 30 days: 575
• Likes: 158
• License: apache-2.0
• Quantized versions: 3
• Fine-tunes: 1
• Tags: zonos2, text-to-speech

Links

• Model: https://huggingface.co/Zyphra/ZONOS2
• Files: https://huggingface.co/Zyphra/ZONOS2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Zyphra/ZONOS2

Then follow the usage in its model card. Nothing in this kit runs by itself.