🦜 VieNeu-TTS v2 Turbo (GPU Edition)

The fastest Bilingual (Vietnamese & English) TTS engine with Instant Zero-Shot Voice Cloning.

---

📖 Model Description

VieNeu-TTS v2 Turbo is the performance-tuned edition of the VieNeu-TTS family. Built on a transformer-based architecture and optimized for minimal latency, it delivers high-fidelity 24 kHz speech synthesis with Instant Voice Cloning capabilities.

This version is designed for GPU-accelerated inference (Standard/Transformers backend), making it ideal for real-time applications, interactive assistants, and creative content generation on platforms like Hugging Face Spaces (ZeroGPU).

✨ Key Features

• 🦜 Instant Voice Cloning: Clone any voice with just 3-5 seconds of reference audio. Truly zero-shot—no reference text required for v2 Turbo!
• 🇻🇳🇺🇸 Bilingual (Code-switching): Seamlessly handles mixed Vietnamese–English sentences in a single utterance.
• 🚀 Extreme Speed: Optimized architecture for ultra-low latency inference on GPUs.
• 🔇 AI Watermarking: Every audio output includes an imperceptible identifier for responsible AI content tracing.
• 🔊 24 kHz High-Fidelity: Studio-quality neural codec output.

---

🚀 Quickstart

Option 1 — Install via vieneu SDK (Recommended)

# Minimal installation (Turbo/CPU Only)
pip install vieneu

# Optional: Pre-built llama-cpp-python for CPU (if building fails)
pip install vieneu --extra-index-url https://pnnbao97.github.io/llama-cpp-python-v0.3.16/cpu/

# Optional: macOS Metal acceleration
pip install vieneu --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/metal/

from vieneu import Vieneu

# Initialize in Turbo mode (Default - Minimal dependencies)
tts = Vieneu()

# 1. Simple synthesis (uses default Southern Male voice 'Xuân Vĩnh')
text = "Hệ thống điện chủ yếu sử dụng alternating current because it is more efficient."
audio = tts.infer(text=text)

# Save to file
tts.save(audio, "output_Xuân Vĩnh.wav")
print("💾 Saved to output_Xuân Vĩnh.wav")

# 2. Using a specific Preset Voice
voices = tts.list_preset_voices()
for desc, voice_id in voices:
    print(f"Voice: {desc} (ID: {voice_id})")

my_voice_id = voices[1][1] if len(voices) > 1 else voices[0][1] # Giọng Phạm Tuyên
voice_data = tts.get_preset_voice(my_voice_id)

audio_custom = tts.infer(text="Tôi đang nói bằng giọng của Bác sĩ Tuyên.", voice=voice_data)

# 3. Save to file
tts.save(audio_custom, "output_Phạm Tuyên.wav")
print("💾 Saved to output_Phạm Tuyên.wav")

🦜 Zero-shot Voice Cloning (SDK)

Clone any voice with only 3-5 seconds of audio using the local Turbo engine:

from vieneu import Vieneu

tts = Vieneu() # Defaults to Turbo mode

# 1. Encode the reference audio (extracts speaker embedding)
# Supported formats: .wav, .mp3, .flac
my_voice = tts.encode_reference("examples/audio_ref/example.wav")

# 2. Synthesize with the cloned voice
# No reference text required for Turbo v2!
audio = tts.infer(
    text="Đây là giọng nói được clone trực tiếp bằng SDK của VieNeu-TTS.", 
    voice=my_voice
)

tts.save(audio, "cloned_voice.wav")

Option 2 — Web UI (Full repo)

git clone https://github.com/pnnbao97/VieNeu-TTS.git
cd VieNeu-TTS
uv sync          # minimal install (Turbo/CPU)
uv run vieneu-web
# → Open http://127.0.0.1:7860

---

🔬 Model Architecture

VieNeu-TTS v2 Turbo utilizes a state-of-the-art two-stage pipeline:

1. Transformer LLM Backbone: A decoder-only transformer that predicts discrete audio tokens from text and speaker embeddings.
2. Neural Codec (VieNeu-Codec): A high-performance VQ-VAE decoder that converts tokens into a 24 kHz waveform with minimal artifacts.

---

📊 Training Data

Trained on a massive multi-speaker dataset comprising over 20,000 hours of high-quality speech:

---

🗺️ Roadmap

• Turbo GPU (Transformers) Engine
• Bilingual (Vietnamese–English) Support
• Zero-shot Voice Cloning
• Mobile SDK (Android / iOS)
• Streaming API Integration

---

🤝 Support & Links

---

📄 License

Released under Apache License 2.0 — permissible for both personal and commercial use.

---

Made with ❤️ for the Vietnamese TTS community by @pnnbao97 and contributors.