vieneu 1.2.6

Advanced on-device Vietnamese TTS with instant voice cloning

🦜 VieNeu-TTS

VieNeu-TTS is an advanced on-device Vietnamese Text-to-Speech (TTS) model with instant voice cloning.

[!TIP] Voice Cloning: All model variants (including GGUF) support instant voice cloning with just 3-5 seconds of reference audio.

This project features two core architectures trained on the VieNeu-TTS-1000h dataset:

• VieNeu-TTS (0.5B): An enhanced model fine-tuned from the NeuTTS Air architecture for maximum stability.
• VieNeu-TTS-0.3B: A specialized model trained from scratch using the VieNeu-TTS-1000h dataset, delivering 2x faster inference and ultra-low latency.

These represent a significant upgrade from the previous VieNeu-TTS-140h with the following improvements:

• Enhanced pronunciation: More accurate and stable Vietnamese pronunciation
• Code-switching support: Seamless transitions between Vietnamese and English
• Better voice cloning: Higher fidelity and speaker consistency
• Real-time synthesis: 24 kHz waveform generation on CPU or GPU
• Multiple model formats: Support for PyTorch, GGUF Q4/Q8 (CPU optimized), and ONNX codec

VieNeu-TTS delivers production-ready speech synthesis fully offline.

Author: Phạm Nguyễn Ngọc Bảo

---

📌 Table of Contents

1. 📦 Using the Python SDK
2. 🐳 Docker & Remote Server
3. 🎯 Custom Models
4. 🛠️ Fine-tuning Guide
5. 🔬 Model Overview
6. 🤝 Support & Contact

---

📦 1. Using the Python SDK (vieneu)

Integrate VieNeu-TTS into your own software projects.

Quick Install

# Option 1: Recommended (Automatically handles hardware acceleration)
uv pip install vieneu

# Option 2: Standard pip (Manual index required for optimization)
# Windows (CPU optimized)
pip install vieneu --extra-index-url https://pnnbao97.github.io/llama-cpp-python-v0.3.16/cpu/

# macOS (Metal GPU accelerated)
pip install vieneu --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/metal/

# Linux / Generic
pip install vieneu

Quick Start (main.py)

from vieneu import Vieneu
import os

# Initialization
tts = Vieneu()

# Standard synthesis (uses default voice)
text = "Xin chào, tôi là VieNeu. Tôi có thể giúp bạn đọc sách, làm chatbot thời gian thực, hoặc thậm chí clone giọng nói của bạn."
audio = tts.infer(text=text)
tts.save(audio, "standard_output.wav")
print("💾 Saved synthesis to: standard_output.wav")

For full implementation details, see examples/main.py.

---

🐳 2. Docker & Remote Server

Deploy VieNeu-TTS as a high-performance API Server (powered by LMDeploy) with a single command.

1. Run with Docker (Recommended)

Requirement: NVIDIA Container Toolkit is required for GPU support.

Start the Server with a Public Tunnel (No port forwarding needed):

docker run --gpus all -p 23333:23333 pnnbao/vieneu-tts:serve --tunnel

• Default: The server loads the VieNeu-TTS model for maximum quality.
• Tunneling: The Docker image includes a built-in bore tunnel. Check the container logs to find your public address (e.g., bore.pub:31631).

2. Using the SDK (Remote Mode)

Once the server is running, you can connect from anywhere (Colab, Web Apps, etc.) without loading heavy models locally:

from vieneu import Vieneu
import os

# Configuration
REMOTE_API_BASE = 'http://your-server-ip:23333/v1'  # Or bore tunnel URL
REMOTE_MODEL_ID = "pnnbao-ump/VieNeu-TTS"

# Initialization (LIGHTWEIGHT - only loads small codec locally)
tts = Vieneu(mode='remote', api_base=REMOTE_API_BASE, model_name=REMOTE_MODEL_ID)
os.makedirs("outputs", exist_ok=True)

# List remote voices
available_voices = tts.list_preset_voices()
for desc, name in available_voices:
    print(f"   - {desc} (ID: {name})")

# Use specific voice (dynamically select second voice)
if available_voices:
    _, my_voice_id = available_voices[1]
    voice_data = tts.get_preset_voice(my_voice_id)
    audio_spec = tts.infer(text="Chào bạn, tôi đang nói bằng giọng của bác sĩ Tuyên.", voice=voice_data)
    tts.save(audio_spec, f"outputs/remote_{my_voice_id}.wav")
    print(f"💾 Saved synthesis to: outputs/remote_{my_voice_id}.wav")

# Standard synthesis (uses default voice)
text_input = "Chế độ remote giúp tích hợp VieNeu vào ứng dụng Web hoặc App cực nhanh mà không cần GPU tại máy khách."
audio = tts.infer(text=text_input)
tts.save(audio, "outputs/remote_output.wav")
print("💾 Saved remote synthesis to: outputs/remote_output.wav")

# Zero-shot voice cloning (encodes audio locally, sends codes to server)
if os.path.exists("examples/audio_ref/example_ngoc_huyen.wav"):
    cloned_audio = tts.infer(
        text="Đây là giọng nói được clone và xử lý thông qua VieNeu Server.",
        ref_audio="examples/audio_ref/example_ngoc_huyen.wav",
        ref_text="Tác phẩm dự thi bảo đảm tính khoa học, tính đảng, tính chiến đấu, tính định hướng."
    )
    tts.save(cloned_audio, "outputs/remote_cloned_output.wav")
    print("💾 Saved remote cloned voice to: outputs/remote_cloned_output.wav")

For full implementation details, see: examples/main_remote.py

Voice Preset Specification (v1.0)

VieNeu-TTS uses the official vieneu.voice.presets specification to define reusable voice assets. Only voices.json files following this spec are guaranteed to be compatible with VieNeu-TTS SDK ≥ v1.x.

3. Advanced Configuration

Customize the server to run specific versions or your own fine-tuned models.

Run the 0.3B Model (Faster):

docker run --gpus all pnnbao/vieneu-tts:serve --model pnnbao-ump/VieNeu-TTS-0.3B --tunnel

Serve a Local Fine-tuned Model: If you have merged a LoRA adapter, mount your output directory to the container:

# Linux / macOS
docker run --gpus all \
  -v $(pwd)/finetune/output:/workspace/models \
  pnnbao/vieneu-tts:serve \
  --model /workspace/models/merged_model --tunnel

For full implementation details, see: main_remote.py

---

🎯 3. Custom Models (LoRA, GGUF, Finetune)

VieNeu-TTS allows you to load custom models directly from HuggingFace or local paths via the Web UI.

👉 See the detailed guide at: docs/CUSTOM_MODEL_USAGE.md

---

🛠️ 4. Fine-tuning Guide

Train VieNeu-TTS on your own voice or custom datasets.

• Simple Workflow: Use the train.py script with optimized LoRA configurations.
• Documentation: Follow the step-by-step guide in finetune/README.md.
• Notebook: Experience it directly on Google Colab via finetune/finetune_VieNeu-TTS.ipynb.

---

🔬 5. Model Overview (Backbones)

Model	Format	Device	Quality	Speed
VieNeu-TTS	PyTorch	GPU/CPU	⭐⭐⭐⭐⭐	Very Fast with lmdeploy
VieNeu-TTS-0.3B	PyTorch	GPU/CPU	⭐⭐⭐⭐	Ultra Fast (2x)
VieNeu-TTS-q8-gguf	GGUF Q8	CPU/GPU	⭐⭐⭐⭐	Fast
VieNeu-TTS-q4-gguf	GGUF Q4	CPU/GPU	⭐⭐⭐	Very Fast
VieNeu-TTS-0.3B-q8-gguf	GGUF Q8	CPU/GPU	⭐⭐⭐⭐	Ultra Fast (1.5x)
VieNeu-TTS-0.3B-q4-gguf	GGUF Q4	CPU/GPU	⭐⭐⭐	Extreme Speed (2x)

🔬 Model Details

• Training Data: VieNeu-TTS-1000h — 443,641 curated Vietnamese samples (Used for all versions).
• Audio Codec: NeuCodec (Torch implementation; ONNX & quantized variants supported).
• Context Window: 2,048 tokens shared by prompt text and speech tokens.
• Output Watermark: Enabled by default.

---

📚 References

• Dataset: VieNeu-TTS-1000h (Hugging Face)
• Model 0.5B: pnnbao-ump/VieNeu-TTS
• Model 0.3B: pnnbao-ump/VieNeu-TTS-0.3B
• LoRA Guide: docs/CUSTOM_MODEL_USAGE.md

---

🤝 6. Support & Contact

• Hugging Face: pnnbao-ump
• Discord: Join our community
• Facebook: Pham Nguyen Ngoc Bao
• Licensing:
  • VieNeu-TTS (0.5B): Apache 2.0 (Free to use).
  • VieNeu-TTS-0.3B: CC BY-NC 4.0 (Non-commercial).
    • ✅ Free: For students, researchers, and non-profit purposes.
    • ⚠️ Commercial/Enterprise: Contact the author for licensing.

---

📑 Citation

@misc{vieneutts2026,
  title        = {VieNeu-TTS: Vietnamese Text-to-Speech with Instant Voice Cloning},
  author       = {Pham Nguyen Ngoc Bao},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/pnnbao-ump/VieNeu-TTS}}
}

---

🙏 Acknowledgements

This project builds upon the NeuTTS Air and NeuCodec architectures. Specifically, the VieNeu-TTS (0.5B) model is fine-tuned from NeuTTS Air, while the VieNeu-TTS-0.3B model is a custom architecture trained from scratch using the VieNeu-TTS-1000h dataset.

---

Made with ❤️ for the Vietnamese TTS community