#VieNeu-TTS: Local Vietnamese Text-to-Speech with Instant Voice Cloning

VieNeu-TTS is a Vietnamese text-to-speech model that runs locally on your CPU, clones voices in seconds, and is open-source under Apache 2.0.

#What it is

VieNeu-TTS is a text-to-speech model built specifically for Vietnamese, designed to run on-device without requiring a GPU. It supports instant voice cloning from just 3 to 5 seconds of reference audio, bilingual Vietnamese-English output with natural code-switching, and real-time inference on CPU.

The model is compact, around 1GB, with a 0.3B quantized variant optimized for CPU performance. Audio output is 24kHz, clear and natural. It is fine-tuned from a 1.5B LLM backbone using a modern TTS architecture, and it ships with a Gradio web UI for quick testing. The training dataset includes roughly 1,000 hours of Vietnamese speech, and the project is released under the Apache 2.0 license.

#Why it matters

• Local, on-device, real-time on CPU. Most high-quality TTS models require GPU inference or cloud API calls. VieNeu-TTS runs on CPU, which means it works on standard laptops and embedded devices without specialized hardware.
• Instant voice cloning with minimal reference. Three to five seconds of audio is enough to clone a voice. This is a remarkably low barrier compared to models that require minutes of clean reference audio or fine-tuning runs.
• Bilingual support with natural code-switching. Vietnamese text frequently mixes in English terms, especially in technical contexts. A TTS model that handles both languages and switches between them naturally solves a real problem for Vietnamese-language applications.
• Privacy-first. Running locally means no audio data leaves your device. This matters for applications in healthcare, education, and personal use where sending voice data to a cloud API is not acceptable.
• Open-source with a real dataset. Apache 2.0 license and 1,000 hours of training data mean you can inspect, modify, and build on this model without vendor lock-in.

#How it works

VieNeu-TTS is fine-tuned from a 1.5B parameter LLM backbone, adapted for speech synthesis. The architecture follows a modern TTS stack that treats text-to-speech as a sequence-to-sequence problem with audio tokens as the output. Voice cloning works by conditioning the model on a short reference clip, which encodes the speaker's voice characteristics. The model then generates speech in that voice from new text input.

The quantized 0.3B variant is specifically optimized for CPU inference, trading some quality for speed and lower memory footprint. The Gradio web UI lets you paste text, upload a reference clip, and generate audio in a browser without writing code. For longer texts, the recommendation is to split input into smaller segments to maintain output quality.

#Caveats

The model is impressive for its size and local-first design, but it has real limitations. For longer passages, output quality can degrade, which is why the project itself recommends splitting text into smaller chunks. This is a workaround, not a feature, and it adds friction to any production pipeline.

The 1,000-hour training dataset is substantial for a community project but small compared to commercial TTS systems trained on tens of thousands of hours. Edge cases in pronunciation, tone, and prosody will exist, especially with uncommon words, names, or mixed-language text.

Voice cloning with only 3 to 5 seconds of reference audio is convenient but inherently limited. The cloned voice will capture broad characteristics but may miss nuance, especially for voices with distinctive speech patterns. Results will vary significantly depending on the quality and cleanliness of the reference clip.

Finally, while CPU inference is a strength, real-time performance depends heavily on your hardware. On older or lower-end CPUs, "real-time" may mean slower than real-time for longer passages.

#Who it's for

VieNeu-TTS is for developers building Vietnamese-language voice applications who need local, privacy-preserving TTS without GPU dependencies. Use cases include audiobook readers, virtual assistants for apps and smart home devices, accessibility tools for visually impaired users, video and podcast narration, and educational applications. If you need to clone a family member's voice for personal content or preserve a voice for sentimental reasons, this is one of the most accessible tools to do it.

The takeaway: if you need Vietnamese TTS that runs locally, clones voices quickly, and does not lock you into a cloud provider, VieNeu-TTS is one of the most practical open-source options available right now. Test it thoroughly against your specific use case before relying on it in production.