ZeroTTS

ZeroTTS

Reads Vietnamese text aloud in real time on a laptop CPU, with streaming output and voice cloning from a few seconds of reference audio. For developers adding Vietnamese speech to apps and media.

Made by zeroweight-ai 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: ZeroTTS is a Vietnamese text-to-speech model that runs through the ONNX library. It turns text into speech on an ordinary CPU, can start playing audio in about 70 milliseconds, and clones a voice from 3 to 30 seconds of reference audio, with no per-speaker training. It also handles English words mixed into Vietnamese text.

Who it suits: It suits developers and producers who need natural Vietnamese speech without a GPU. Clone only voices you have the right to use.

Industry | Who uses it | How they use it | Example
Media | News and podcast producers | Narrate Vietnamese scripts, including English words inside them | A news script read aloud in a chosen voice
Software | App developers | Add streaming Vietnamese speech to an app | A voice assistant that starts speaking almost at once
Education | Course makers and teachers | Turn written lessons into audio | A Vietnamese lesson narrated for listeners

You can skip it if you need a text-to-speech model built mainly for other languages.

Where to use it:

• Python scripts and apps on Windows, macOS and Linux, on a CPU with no GPU
• Live playback through the sounddevice package, which the streaming example uses
• A browser demo, which the author's code page offers

Use case | Who | What they get back
Reading text aloud | Developers | A WAV file of the spoken text
Live speech | App builders | Audio chunks you can play as they arrive
Voice cloning | Producers | Speech in a voice copied from a short reference clip

How to set it up: You need these first:

• Python on Windows, macOS or Linux
• The zerotts package
• The sounddevice package, only for the streaming playback example

1. Install the zerotts package with pip.
2. Optionally download the repository's files with Hugging Face's client.
3. Load the model in Python with from_pretrained and the model name.

pip install zerotts

pip install -U huggingface_hub
hf download zeroweight-ai/ZeroTTS

How to use it: You load the model, pass it text and a voice name, then save the audio. For live playback, the card also shows a streaming call that returns audio in chunks.

from zerotts import ZeroTTS

tts = ZeroTTS.from_pretrained("zeroweight-ai/ZeroTTS")
audio = tts.synthesize("Xin chào các bạn, mình là ZeroTTS.", voice="maichi")
tts.save_audio(audio, "out.wav")

Tool | What it does
ZeroTTS.from_pretrained | Loads the model by its name
synthesize | Turns text into audio with the voice you name
synthesize_stream | Yields audio chunks as they are made, for live playback
save_audio | Writes the audio to a WAV file

Example: You ask it to say a Vietnamese greeting, "Xin chào các bạn, mình là ZeroTTS.", in the voice named maichi. The code saves out.wav, a short recording of that line spoken aloud.

For newcomers and experts: ZeroTTS reads text aloud. A good first try is installing the package, running the short script above, and playing out.wav. If it sounds right, try changing the sentence.

For readers who know the field, a voice is a small latent array, so a new voice needs reference audio and no fine-tuning. The model has 202M parameters, and the card gives a real-time factor of 0.50 on CPU. Its numbers come from the author's own benchmark dataset, where it scores 1.03 percent word error against 4.13 for the next system.

Install zerotts with pip
Load the model with from_pretrained
Pass text and a voice name
Get audio in one piece or in chunks
Save it or play it live

Good to know:

• Language: built for Vietnamese. The card also shows an English output from a Vietnamese reference clip.
• Speed: the card reports about 70 ms to first audio and a real-time factor of 0.50 on CPU.
• Benchmarks: the author measured them on its own ZeroBench-TTS set, so treat the comparison with care.

Title: Word error rate on the author's benchmark, lower is better
ZeroTTS: 1.03%
OmniVoice: 4.13%
XTTS-v2-vietnamse: 16.42%
viXTTS: 18.40%

• License: MIT.
• Cost to run: none named. It runs on a CPU, and no GPU is needed.
• Data it reaches: the text you give it and any reference audio, on your own machine. Installing and loading fetch files over the web.
• Not for: none named, though cloning someone's voice without their right to it is not a use to pursue.

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: onnx
• Downloads in the last 30 days: 8,091
• Likes: 54
• License: mit
• Languages: vi
• Tags: onnx, text-to-speech, vietnamese, onnxruntime, zero-shot, speech-synthesis, voice-cloning, vietnamese-tts, tieng-viet, model-index

Links

• Model: https://huggingface.co/zeroweight-ai/ZeroTTS
• Files: https://huggingface.co/zeroweight-ai/ZeroTTS/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download zeroweight-ai/ZeroTTS

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