pnnbao-ump
/

VieNeu-TTS-v3-Turbo

 ---
 license: apache-2.0
+datasets:
+- pnnbao-ump/VieNeu-TTS-10k-ENVI
+language:
+- vi
+- en
+pipeline_tag: text-to-speech
+tags:
+- voice-cloning
+- code-switching
+- podcast
+- emotion-control
+- 48khz
 ---
+# 🦜 VieNeu-TTS v3 Turbo
+[![GitHub](https://img.shields.io/badge/GitHub-Repository-blue)](https://github.com/pnnbao97/VieNeu-TTS)
+[![Model](https://img.shields.io/badge/Hugging%20Face-Model-yellow)](https://huggingface.co/pnnbao-ump/VieNeu-TTS-v3-Turbo)
+[![Discord](https://img.shields.io/badge/Discord-Join%20Us-5865F2?logo=discord&logoColor=white)](https://discord.gg/yJt8kzjzWZ)
+## Overview
+<video controls src="https://cdn-uploads.huggingface.co/production/uploads/68b923a86c86c127a1975eda/fY6_ONxsTJ02WgNmeFpfL.mp4" width="100%"></video>
+**VieNeu-TTS v3 Turbo** is the next generation of Vietnamese TTS — **48 kHz** high-fidelity speech, **instant voice cloning**, built-in **multi-speaker default voices**, inline **emotion cues**, and seamless **bilingual (En–Vi) code-switching**. It is a pure-PyTorch engine running on both **GPU and CPU**, using the **MOSS-Audio-Tokenizer-Nano** codec.
+> [!NOTE]
+> **Early access.** v3 Turbo is released for **preview**. It is fast and natural, but some features (notably the emotion cues) are still **experimental**. The full v3 release is coming in the next few weeks.
+> [!IMPORTANT]
+> **What's new in v3:**
+> - **48 kHz audio** — a big jump in fidelity over v2 (24 kHz).
+> - **Built-in default voices** — each default speaker is addressed by a dedicated speaker token + fixed reference, so the voice is **stable and consistent** with no reference clip needed.
+> - **Emotion / non-verbal cues** *(experimental)* — drop `[cười]`, `[thở dài]`, `[hắng giọng]` straight into your text.
+> - **Batched generation** — synthesize many chunks at once (batch size up to 32), including a **multi-speaker conversation** mode that batches the whole script regardless of speaker.
+> - **Instant Voice Cloning** — still clones a voice from just **3–5 seconds** of audio (cloning is available from v3 onward; v1/v2 do not support it).
+## 🏗️ Architecture & Credits
+The **VieNeu-TTS v3 Turbo architecture is an original design by the author, Phạm Nguyễn Ngọc Bảo**, and is **trained from scratch** on ~10,000 hours of English–Vietnamese speech — it is **not** a fine-tune, distillation, or adaptation of any existing TTS model.
+- **Model architecture & training:** designed and trained from scratch by **Phạm Nguyễn Ngọc Bảo** — https://github.com/pnnbao97
+- **Audio codec:** [MOSS-Audio-Tokenizer-Nano](https://huggingface.co/OpenMOSS-Team/MOSS-Audio-Tokenizer-Nano) (OpenMOSS-Team) — 48 kHz neural audio codec.
+- **Phonemizer:** [sea-g2p](https://github.com/pnnbao97/sea-g2p) — fast Vietnamese/English grapheme-to-phoneme, also by the author.
+Tác giả: **Phạm Nguyễn Ngọc Bảo**
+## ☕ Support This Project
+Training high-quality TTS models requires significant GPU resources. If you find this model useful, please consider supporting the development:
+[![Buy Me a Coffee](https://img.shields.io/badge/Buy%20Me%20a%20Coffee-Support-orange?logo=buy-me-a-coffee)](https://buymeacoffee.com/pnnbao)
+---
+## 🔥 Quick Start (Web UI)
+```bash
+git clone https://github.com/pnnbao97/VieNeu-TTS.git
+cd VieNeu-TTS
+# Install uv (if you haven't)
+# Windows: powershell -c "irm https://astral.sh/uv/install.ps1 | iex"
+# Linux/macOS: curl -LsSf https://astral.sh/uv/install.sh | sh
+# Install dependencies & Run
+uv sync --group gpu
+uv run vieneu-web
+```
+In the Web UI, pick **"VieNeu-TTS-v3-Turbo (Thử nghiệm)"** as the backbone. You get a **Default voice** tab, a **Voice Cloning** tab, and a **Conversation** tab (batched multi-speaker podcasts).
+---
+## 📦 Using Python SDK (vieneu)
+```bash
+# Linux / MacOS
+pip install vieneu
+# Windows (avoid llama-cpp build errors)
+pip install vieneu --extra-index-url https://pnnbao97.github.io/llama-cpp-python-v0.3.16/cpu/
+```
+### Full Features Guide
+```python
+from vieneu import Vieneu
+# Load v3 Turbo (48 kHz) together with the built-in default voices.
+tts = Vieneu(mode="v3turbo")
+# 1. Simple synthesis with the default voice (Ngọc Lan)
+text = "Xin chào, đây là VieNeu-TTS phiên bản ba Turbo. Giọng nói tự nhiên, mượt mà như người thật."
+audio = tts.infer(text)
+tts.save(audio, "output.wav")          # 48 kHz
+print("💾 Saved to output.wav")
+# 2. Pick a built-in default voice by name (no reference needed)
+for label, voice_id in tts.list_preset_voices():
+    print(f"Voice: {label}  (ID: {voice_id})")
+audio = tts.infer("Mình là Xuân Vĩnh, rất vui được gặp các bạn.", voice="Xuân Vĩnh")
+tts.save(audio, "output_xuanvinh.wav")
+# 3. Emotion / non-verbal cues — EXPERIMENTAL
+#    [cười] = chuckle, [thở dài] = sigh, [hắng giọng] = clear throat
+audio = tts.infer("Trời ơi nghe hay quá đi [cười]. Để mình nói tiếp nhé [hắng giọng].",
+                  voice="Ngọc Linh")
+tts.save(audio, "output_emotion.wav")
+# 4. Instant voice cloning from a 3–5s reference clip
+audio = tts.infer("Đây là giọng nói được nhân bản tức thì.", ref_audio="my_voice.wav")
+tts.save(audio, "output_clone.wav")
+# 5. Low-latency streaming
+for chunk in tts.infer_stream("Văn bản dài sẽ được phát theo từng đoạn.", voice="Ngọc Lan"):
+    ...  # play / send each chunk as it is produced
+```
+> [!TIP]
+> A **temperature around 0.8** gives the most stable result for v3 Turbo. Higher values add expressiveness but can be less stable.
+---
+## 🎭 Default Voices
+Built-in voices — call them by name via `voice="<name>"`, no reference audio required.
+| Voice | Gender | Style |
+|---|---|---|
+| **Ngọc Lan** *(default)* | Female | Soft / gentle |
+| Ngọc Linh | Female | Bright |
+| Trúc Ly | Female | Youthful |
+| Mỹ Duyên | Female | Smooth |
+| Xuân Vĩnh | Male | Upbeat |
+| Thái Sơn | Male | Firm |
+| Gia Bảo | Male | Smooth |
+| Đức Trí | Male | Clear |
+| Trọng Hữu | Male | Knowledgeable |
+| Bình An | Male | Even / calm |
+For any other voice, use **Voice Cloning** with a short reference clip (`ref_audio="..."`).
+---
+## 🔬 Model Variants
+| Model                 | Format  | Device  | Sample Rate | Quality    | Features                                  |
+| --------------------- | ------- | ------- | ----------- | ---------- | ----------------------------------------- |
+| **VieNeu-TTS-v3-Turbo** | PyTorch | GPU/CPU | 48 kHz      | ⭐⭐⭐⭐⭐ | **Default voices, Cloning, Emotion cues** |
+| VieNeu-TTS-v2         | PyTorch | GPU/CPU | 24 kHz      | ⭐⭐⭐⭐⭐ | Podcast, En-Vi code-switching             |
+| VieNeu-TTS-v2 (GGUF)  | GGUF Q4 | CPU     | 24 kHz      | ⭐⭐⭐⭐   | Fastest on CPU, Podcast                   |
+| VieNeu-TTS-v1         | PyTorch | GPU     | 24 kHz      | ⭐⭐⭐⭐   | Stable (Vi only)                          |
+---
+## 📑 Citation
+```bibtex
+@misc{vieneutts2026,
+  title        = {VieNeu-TTS v3 Turbo: 48kHz Vietnamese Text-to-Speech with Instant Voice Cloning and Emotion Control},
+  author       = {Pham Nguyen Ngoc Bao},
+  year         = {2026},
+  publisher    = {Hugging Face},
+  howpublished = {\url{https://huggingface.co/pnnbao-ump/VieNeu-TTS-v3-Turbo}}
+}
+```
+---
+**Made with ❤️ for the Vietnamese TTS community**