VieNeu-TTS v3 Turbo

VieNeu-TTS v3 Turbo

Speaks Vietnamese and English at 48 kHz with 23 preset voices from three regions, instant voice cloning and real-time streaming. A small 131M text-to-speech model that also runs on a plain CPU.

Made by pnnbao-ump 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: VieNeu-TTS v3 Turbo is a text-to-speech model for Vietnamese and English from pnnbao-ump, with 131M parameters. It writes 48 kHz audio and comes with 23 preset voices from the North, Centre and South of Vietnam. It can also copy a voice from a short sample, stream speech in real time and mix both languages in one sentence. The author says it was designed and trained from scratch on roughly 10,000 hours of English and Vietnamese speech.

Who it suits: Developers and creators who build Vietnamese voice apps, narration, voice bots and reading tools, and who want a small model that works on a normal computer.

Industry | Who uses it | How they use it | Example
Media and video | Creators | Narrate scripts in Vietnamese | A voice-over in a Southern accent
Customer service | Voice bot builders | Stream spoken replies to callers | A support bot that answers in Vietnamese
Education | Language teachers | Read lessons in both languages | A bilingual reading with English terms
App development | Developers | Add a speech endpoint to a product | A service that works with OpenAI client tools

If you need languages other than Vietnamese and English, you can skip it, because the card lists only those two.

Where to use it:

• Python, through the vieneu package
• A browser interface that the project starts on your own computer
• A streaming service with an OpenAI-compatible interface, which the card says works with common voice-agent tools

Use case | Who | What they get back
Default voices | Everyone | Speech in one of 23 preset voices
Voice cloning | Producers | Speech that copies a short sample of a voice
Streaming | Voice bot builders | Audio that starts within a fraction of a second

How to set it up:

• Python and pip
• About 262 MB of disk space in BF16, or about 65 MB at 4-bit (an estimate)
• No GPU needed, since the plain install runs on CPU through ONNX Runtime
• An NVIDIA GPU with CUDA 12.8 or later, only if you want the faster batched and streaming mode

1. Install the vieneu package with pip.
2. Optionally download the model files yourself.
3. Use the package from Python, or start the project's web interface from its repository.

pip install vieneu
hf download pnnbao-ump/VieNeu-TTS-v3-Turbo

How to use it: Install the package and call it from Python. The card also offers a web interface, started with the command below from inside the project's repository folder. It has a tab for the default voices, a tab for cloning and a tab for multi-speaker conversations. The part of the card shown here has no full code sample, so follow its guide.

uv run vieneu-web

Example: You open the web interface, type a Vietnamese sentence with an English product name in the middle and pick one of the preset voices. The model reads the sentence in that voice and saves a 48 kHz audio file.

For newcomers and experts: If you are new to this, begin with the web interface and a preset voice, then try cloning with a clip you own. Only clone a voice that is yours or that you have the right to use.

On a CPU the model runs through ONNX Runtime without PyTorch, and an optional int8 mode makes the CPU path faster. On a CUDA machine the package switches to a PyTorch engine that batches requests and shares one scheduler between streams. The card says one RTX 3060 can serve 16 streams at once. It builds on a separate 48 kHz audio codec and a Vietnamese and English phonemizer.

You type Vietnamese or mixed text
You pick a preset voice or add a sample
The model generates speech
You save or stream 48 kHz audio

Good to know:

• Languages: Vietnamese and English, including mixed sentences
• Voice cloning: use only voices you have the right to use
• Variants: the Hugging Face record lists 3 quantized and 1 fine-tuned version made by others

Title: Time before the first audio, in milliseconds
GPU streaming: 115 ms
CPU with int8: 140 ms
CPU with fp32: 300 ms

• License: apache-2.0
• Cost to run: no API key or paid service is named, only your own computer, and a GPU is optional
• Data it reaches: downloading fetches model files, and the model reads your text and any voice sample you give it
• Not for: cloning a voice without the right to use it, or speech in other languages

On Hugging Face, 13 November 2026

• Task: text to speech
• Parameters: 131M
• Weights: about 262 MB in BF16; about 65 MB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 2,151,800
• Likes: 109
• License: apache-2.0
• Languages: vi, en
• Quantized versions: 3
• Fine-tunes: 2
• Tags: onnx, safetensors, gguf, vieneu_v3_turbo, voice-cloning, code-switching, podcast, emotion-control, 48khz, text-to-speech

Links

• Model: https://huggingface.co/pnnbao-ump/VieNeu-TTS-v3-Turbo
• Files: https://huggingface.co/pnnbao-ump/VieNeu-TTS-v3-Turbo/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download pnnbao-ump/VieNeu-TTS-v3-Turbo

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