VieNeu-TTS v3

A 131M text-to-speech model for Vietnamese and English with 48 kHz output, 23 voices, cloning and CPU inference. Apache 2.0.

In short. VieNeu-TTS reads Vietnamese text aloud, and English too, in a natural voice that you can pick from 23 ready-made ones or copy from a short recording. It is small enough to run on an ordinary computer's processor, and everything, including the voices, is free to use in a business.

Summary

VieNeu-TTS is a Vietnamese text-to-speech project by one independent developer, Phạm Nguyễn Ngọc Bảo. Its v3 Turbo model first appeared on Hugging Face on 5 June 2026. It has 131M parameters, an LLM-style autoregressive backbone, and the MOSS-Audio-Tokenizer-Nano codec from OpenMOSS. The card says it was trained from scratch on about 10,000 hours of English and Vietnamese speech, and that it is not a fine-tune or distillation of another TTS model. It writes 48 kHz speech with 23 preset voices from the North, Central and South regions, clones a voice from a 3 to 8 second clip, switches between Vietnamese and English inside one text, and accepts experimental cues such as [cười] for a laugh.

You use it through the vieneu Python SDK (3.7.1 on the card) or an OpenAI-compatible speech server. On a CPU the SDK is torch-free and runs ONNX Runtime, and the card gives first audio in about 300 ms in fp32 and 140 ms in int8. On an NVIDIA GPU it batches requests, and the card reports first audio in about 115 ms and 16 concurrent streams under 200 ms on one RTX 3060. v3 Nano is a separate preview design with 48M parameters, flow matching and 24 kHz output. It loads in about 3 seconds and runs about 3 times faster than Turbo on a CPU, but the card says its quality is noticeably lower and its English is weak. These are the developer's own numbers, and the cards publish no listening or error-rate benchmark, so judge the voices by ear.

The weights, the ONNX exports and the bundled preset voices are all under Apache 2.0. The card says the speakers behind the presets consented to AI training and synthetic speech, and that generated audio may be used commercially. The training data is not disclosed, and the 10,000-hour dataset is gated. The consent covers only the presets, so cloning someone else's voice needs that person's permission. In the 30 days to 11 November 2026, Hugging Face counted about 2.1 million downloads for Turbo, a counter that also includes automated installs, and about 108,000 for Nano.

Good for

  • Vietnamese narration, audiobooks and voice assistants, with regional North, Central and South voices.
  • Running text to speech on a CPU or a small server, including streaming for many listeners on one mid-range GPU.
  • Vietnamese texts that mix in English words or whole sentences.
  • Commercial products, because the code, weights and preset voices share one Apache 2.0 license.

Watch out for

  • It supports Vietnamese and English only, and Nano is weak at English.
  • It is a single-author project with undisclosed training data and no published quality benchmark.
  • The inline emotion cues are marked experimental, and the preset voice list changes between SDK versions, so pin the version you test.
  • v3 Nano is a preview with lower quality, 24 kHz output and a preset list that may still change.
  • Cloning a real person's voice is your responsibility, and the consent statement covers only the bundled presets.

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of VieNeu-TTS v3. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.

Variants Parameters Download size Memory needed Fits on
v3 Nano (preview) 48M 0.28 GB 0.322 GB Laptop or phone
v3 Turbo BF16 131M 0.26 GB 0.299 GB Laptop or phone

Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is VieNeu-TTS v3?
VieNeu-TTS reads Vietnamese text aloud, and English too, in a natural voice that you can pick from 23 ready-made ones or copy from a short recording. It is small enough to run on an ordinary computer's processor, and everything, including the voices, is free to use in a business.
Can I use VieNeu-TTS v3 commercially?
Yes. VieNeu-TTS v3 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is VieNeu-TTS v3 open source?
VieNeu-TTS v3 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run VieNeu-TTS v3 on my own computer?
Yes. The smallest variant, v3 Nano (preview), needs about 0.322 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was VieNeu-TTS v3 released?
VieNeu-TTS v3 was released by Phạm Nguyễn Ngọc Bảo (VieNeu-TTS) on June 5, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Kokoro hexgrad 82M – Apache 2.0 (Permissive) 0.38 GB Laptop or phone
Supertonic 3 Supertone 99M – OpenRAIL-M (Conditional) 0.46 GB Laptop or phone
Pocket TTS Kyutai 18M – 100M – CC BY 4.0 (Permissive) 0.127 GB Laptop or phone
KittenTTS 2 KittenML (Stellon Labs) 1.7B – Stellon Labs Community (Conditional) 6 GB Laptop or phone