V-TTS

Vietnamese Text-to-Speech with Multi-Speaker TTS and Zero-Shot Voice Cloning.

The lightest Vietnamese zero-shot voice cloning model — only 74.8M parameters, runs entirely on CPU, 3-4x faster than realtime. No GPU required.

📅 Cập nhật: Tháng 4/2026 — Đã phát hành phiên bản Windows App (.exe), không cần cài đặt Python hay chạy source code phức tạp!

📥 Download Windows App (Khuyến nghị)

Dành cho người dùng Windows — Tải file .exe, chạy trực tiếp, không cần cài đặt Python, không lo lỗi source code!

Phiên bản	Hệ điều hành	Download
v1.0.5 (Mới nhất)	Windows 10/11 (64-bit)	📥 Tải VTTS.exe (394MB)

✨ Tính năng Windows App:

✅ Không cần cài đặt Python — Chạy trực tiếp file .exe
✅ Không lo lỗi source code — Đã đóng gói sẵn toàn bộ dependencies
✅ Giao diện đẹp — Dark mode hiện đại, dễ sử dụng
✅ 5 giọng đọc có sẵn — NF, SF, NM1, SM, NM2 (Bắc/Nam, Nam/Nữ)
✅ Xử lý hàng loạt — Tạo nhiều audio từ file .txt/.srt
✅ Tốc độ tùy chỉnh — Điều chỉnh từ 0.5x đến 2.0x
✅ Nghe thử & Lưu file — Preview và export audio dễ dàng

Ảnh chụp ứng dụng:

🚀 Hướng dẫn sử dụng:

Tải file VTTS.exe từ Releases
Chạy file .exe (lần đầu sẽ tự tải model từ HuggingFace — cần kết nối Internet)
Nhập văn bản tiếng Việt → Chọn giọng đọc → Nhấn "Tạo Audio"
Nghe thử hoặc lưu file .wav

Lưu ý: File .exe có dung lượng ~394MB do đã bao gồm toàn bộ PyTorch và dependencies. Lần chạy đầu tiên sẽ tải model (~100MB) từ HuggingFace.

🐳 Chạy bằng Docker (Windows / Mac / Linux)

Dành cho mọi hệ điều hành — Chạy trong Docker, không cần cài đặt Python, cách ly hoàn toàn!

Yêu cầu:

Docker Desktop đã cài đặt

Cách chạy:

Cách 1: Dùng docker-compose (Khuyến nghị)

git clone https://github.com/tronghieuit/v-tts.git
cd v-tts
docker-compose up -d

Cách 2: Dùng Docker trực tiếp

docker build -t v-tts .
docker run -d -p 7860:7860 --name v-tts v-tts

Mở trình duyệt: http://localhost:7860

✨ Tính năng Docker:

✅ Đa nền tảng — Windows, macOS, Linux
✅ Cách ly hoàn toàn — Không ảnh hưởng hệ thống
✅ Gradio Web UI — Giao diện web đẹp, dễ dùng
✅ Auto-download model — Tự tải model lần đầu
✅ 5 giọng đọc — NF, SF, NM1, SM, NM2

Lưu ý: Docker image ~2-3GB do bao gồm PyTorch và dependencies. Lần chạy đầu tiên sẽ tải model từ HuggingFace.

Highlights

🪶 Ultra-lightweight: 74.8M params — the lightest Vietnamese zero-shot voice clone model
⚡ CPU-only: RTF ~0.24 on CPU (4x faster than realtime), no GPU needed
🎯 Zero-shot: Clone any voice from 3-10s of audio, no fine-tuning
🎨 Prosody Transfer: Transfer intonation, rhythm, emotion from reference voice
🇻🇳 Vietnamese-native: Dedicated Vietnamese phonemizer with Northern/Southern support
📦 Multi-speaker TTS: 5 built-in Vietnamese voices (Northern/Southern, Male/Female)
🔌 Simple API: pip install and use with 2 lines of code

🚀 Live Demo — Try it now!

Demo	Link
🎙️ Zero-Shot Voice Cloning	▶️ huggingface.co/spaces/v-tts/v-zeroshot-voice-cloning
🔊 Multi-Speaker TTS	▶️ huggingface.co/spaces/v-tts/v-vietnamese-tts

Clone any voice from 3-10 seconds of audio. No GPU required. Try it directly in your browser!

▶️ Watch Zero-Shot Demo Video

🎧 Zero-Shot Voice Cloning Examples

Same text, cloned with 6 different reference voices:

Reference Voice	Description	Cloned Audio
Thu Hà	Soft female voice	▶️ example_thu_ha.wav
Minh Đức	Deep male voice	▶️ example_minh_duc.wav
Thanh Tâm	Young female voice	▶️ example_thanh_tam.wav
Quang Huy	Young male voice	▶️ example_quang_huy.wav
Ngọc Ánh	Professional female	▶️ example_ngoc_anh.wav
Hoàng Nam	Strong male voice	▶️ example_hoang_nam.wav

Multi-Speaker TTS Examples

Speaker	Region	Gender	Audio
NF	Northern	Female	▶️ example_NF.wav
SF	Southern	Female	▶️ example_SF.wav
NM1	Northern	Male	▶️ example_NM1.wav
SM	Southern	Male	▶️ example_SM.wav
NM2	Northern	Male	▶️ example_NM2.wav

Clone the repo and listen to files in examples/ for the best audio quality.

⚡ Performance — CPU is all you need

Zero-Shot Voice Cloning (74.8M params)

Component	Parameters	Purpose
Synthesizer	56.45M	Voice synthesis
Speaker Encoder	8.03M	Voice identity extraction
Style Encoder	7.80M	Prosody/style extraction
Prosody Predictor	2.52M	F0/energy prediction
Total	74.80M	~285 MB (FP32)

Zero-Shot CPU Benchmark

Runs entirely on CPU — no GPU needed!

Input Length	Inference Time	Audio Length	RTF	Speed
Short (16 chars)	506ms	1.07s	0.475	2.1x realtime
Medium (60 chars)	1,030ms	3.61s	0.286	3.5x realtime
Long (156 chars)	2,229ms	9.45s	0.236	4.2x realtime

Multi-Speaker TTS (57.97M params)

Mode	Short	Medium	Long
CPU (i5-14500)	392ms / RTF 0.54	854ms / RTF 0.38	1,653ms / RTF 0.34
CUDA (RTX 4060 Ti)	45ms / RTF 0.06	52ms / RTF 0.02	69ms / RTF 0.01

RTF (Real-Time Factor): Processing time / audio duration. RTF < 1 = faster than realtime.

Installation

# From Git
pip install git+https://github.com/tronghieuit/v-tts.git

# From Source
git clone https://github.com/tronghieuit/v-tts.git
cd v-tts
pip install -e .

Requirements

Python 3.8+
PyTorch 2.0+
No GPU required (CUDA optional for multi-speaker acceleration)
Linux recommended for best phonemization quality

Quick Start

Multi-Speaker TTS (2 lines)

from v_tts import TTS

tts = TTS()  # Auto-downloads model from Hugging Face
tts.speak("Xin chào các bạn", speaker="NF", output_path="hello.wav")

# Get audio array
audio, sr = tts.synthesize("Xin chào các bạn", speaker="NM1")

# Available speakers: NF, SF, NM1, SM, NM2
print(tts.list_speakers())

Zero-Shot Voice Cloning (2 lines)

from v_tts import ZeroShotTTS

tts = ZeroShotTTS()  # Auto-downloads model, CPU by default
tts.clone_voice(
    text="Xin chào, tôi là giọng nói được nhân bản",
    reference_audio="your_voice.wav",
    output_path="output.wav"
)

# Or get audio array
audio, sr = tts.synthesize(
    text="Đây là văn bản tiếng Việt",
    reference_audio="references/thu_ha.wav"
)

Command Line

# Zero-shot voice cloning
python infer_zeroshot.py \
  --reference references/thu_ha.wav \
  --text "Buổi sáng ở thành phố bắt đầu bằng những âm thanh quen thuộc" \
  --output cloned_voice.wav

# Use your own voice
python infer_zeroshot.py \
  --reference your_voice.wav \
  --text "Văn bản tiếng Việt của bạn" \
  --output output.wav --cpu

# Multi-speaker TTS
python infer.py --text "Xin chào các bạn" --speaker NF --output hello.wav
python infer.py --interactive

Gradio Demos

# Zero-shot voice cloning demo
python app_zeroshot.py
# Open: http://localhost:7860

# Multi-speaker TTS demo
python app.py

🎤 Zero-Shot Voice Cloning — Details

Architecture

Reference Audio
    ├─→ Speaker Encoder → Speaker Embedding (512-dim)
    └─→ Mel Extraction → Style Encoder → Prosody Embedding (128-dim)

Input Text → Vietnamese Phonemizer → Text Encoder → Text Representation

[Speaker Emb + Prosody Emb + Text Repr] → Voice Generator → Output Audio

Components

Speaker Encoder (512-dim) — Extracts speaker identity from raw audio waveform. L2-normalized for generalization.
Style Encoder (128-dim) — Extracts prosody/style from mel spectrograms (80 channels). Captures rhythm, emotion, speaking style.
Prosody Predictor — Predicts F0 (pitch) and energy contours from text + style embedding.
Voice Generator — Neural vocoder with style-dependent normalization (FiLM conditioning for F0/energy).

Reference Audio Requirements

For best results:

✅ Good	❌ Bad
3-10 seconds duration	< 2s (poor representation)
Clean, clear speech	Noisy background
Single speaker	Multiple speakers
Neutral emotion	Extreme emotion/shouting
Any language	Music or sound effects

# Prepare reference audio
ffmpeg -i full_audio.wav -ss 00:00:10 -t 5 -ar 24000 reference.wav
ffmpeg -i input.mp3 -ar 24000 -ac 1 reference.wav

6 Built-in Reference Voices

Available in references/:

Voice	File	Description
Thu Hà	`thu_ha.wav`	Soft, warm female voice
Minh Đức	`minh_duc.wav`	Deep, composed male voice
Thanh Tâm	`thanh_tam.wav`	Young, bright female voice
Quang Huy	`quang_huy.wav`	Young, energetic male voice
Ngọc Ánh	`ngoc_anh.wav`	Professional female anchor
Hoàng Nam	`hoang_nam.wav`	Strong, deep male voice

Upload your own reference audio (3-10 seconds, clear speech) to clone any voice.

Advanced Usage

Speaker Interpolation

# Blend two voices
spk_emb_1 = tts.extract_embeddings("voice_1.wav")[0]
spk_emb_2 = tts.extract_embeddings("voice_2.wav")[0]

alpha = 0.5  # 50-50 blend
spk_emb_mixed = alpha * spk_emb_1 + (1 - alpha) * spk_emb_2

Limitations

Vietnamese synthesis works best (model supports multiple languages but optimized for Vietnamese)
Cannot improve upon reference audio quality
Very unique voices may not be perfectly replicated
Not optimized for real-time streaming (yet)

🔊 Multi-Speaker TTS — Details

5 Vietnamese Voices

Speaker	Region	Gender	Code
NF	Northern (Miền Bắc)	Female	`NF`
SF	Southern (Miền Nam)	Female	`SF`
NM1	Northern (Miền Bắc)	Male	`NM1`
SM	Southern (Miền Nam)	Male	`SM`
NM2	Northern (Miền Bắc)	Male	`NM2`

Parameters

Parameter	Default	Description
`speed`	1.0	< 1.0 = faster, > 1.0 = slower
`noise_scale`	0.667	Voice variability
`noise_scale_w`	0.8	Duration variability
`sdp_ratio`	0.0	0 = deterministic, 1 = stochastic

Model Auto-Download

Models are automatically downloaded from Hugging Face and cached:

Windows: %LOCALAPPDATA%\v_tts\models\
Linux/Mac: ~/.cache/v_tts/models/

Project Structure

v-tts/
├── v_tts/           # pip install package
│   ├── tts.py            # Multi-speaker TTS API
│   └── zeroshot.py       # Zero-shot voice cloning API
├── src/
│   ├── models/           # All model architectures
│   │   ├── synthesizer_zeroshot.py  # Zero-shot synthesizer
│   │   ├── synthesizer.py          # Multi-speaker synthesizer
│   │   ├── encoders.py             # Speaker/Style/Prosody encoders
│   │   └── adain.py                # Style conditioning module
│   ├── nn/               # Neural network components + mel processing
│   ├── text/             # Text/phoneme processing
│   └── vietnamese/       # Vietnamese text normalization + phonemizer
├── pretrained/
│   ├── zeroshot/         # Zero-shot checkpoint + config
│   ├── hasp/             # Speaker encoder weights
│   ├── onnx/             # ONNX export models
│   └── config.json       # Multi-speaker config
├── references/           # 6 reference voices for zero-shot
├── examples/
│   ├── zeroshot/         # Zero-shot cloned audio samples
│   └── *.wav             # Multi-speaker audio samples
├── deployments/          # Edge / Web / Android deployments
├── infer_zeroshot.py     # Zero-shot CLI inference
├── infer.py              # Multi-speaker CLI inference
├── app_zeroshot.py       # Gradio demo (zero-shot)
└── app.py                # Gradio demo (multi-speaker)

📱 Deployment Options

Platform	Technology	Size	Offline
HuggingFace Spaces	Gradio	Cloud	No
Edge	ONNX Runtime	~165MB	Yes
Web	ONNX Runtime Web	~165MB	Yes
Android	ONNX Runtime Mobile	~185MB	Yes

See deployments/ for detailed guides.

Citation

@software{v_tts,
  title = {V-TTS with Zero-Shot Voice Cloning},
  author = {V-TTS},
  year = {2026},
  url = {https://github.com/tronghieuit/v-tts}
}

License

CC BY-NC 4.0 — Non-commercial use only. Commercial use requires written permission.

Acknowledgments

V-TTS for model training and development
Vietnamese phonemization community

Name		Name	Last commit message	Last commit date
assets		assets
deployments		deployments
dist		dist
examples		examples
pyi_hooks		pyi_hooks
src		src
v_tts		v_tts
.dockerignore		.dockerignore
.gitattributes		.gitattributes
.gitignore		.gitignore
Dockerfile		Dockerfile
LICENSE		LICENSE
README.md		README.md
VTTS.spec		VTTS.spec
benchmark.py		benchmark.py
build_exe.bat		build_exe.bat
demo_gradio.py		demo_gradio.py
docker-compose.yml		docker-compose.yml
generate_examples.py		generate_examples.py
gui_app_modern.py		gui_app_modern.py
infer.py		infer.py
requirements.txt		requirements.txt
setup.py		setup.py
validate_docker.py		validate_docker.py

Latest commit

History

Folders and files

Repository files navigation