SNAP Voice (snap_voice)

🏠 SNAP Project Homepage
🌐 SNAP Main Hub (snap) | ⚡ SNAP C++ SDK (snap_cpp) | ✉️ Contact

The snap_voice project aims to deliver high-quality end-to-end speech synthesis by integrating the high-performance SNAP C++ Engine (snap_cpp) with various backend Text-to-Speech (TTS) engines.

As our first milestone, we integrated the MeloTTS backend with the SNAP C++ Frontend, achieving significantly improved speech synthesis quality, precise Inverse Text Normalization (ITN), and contextual Phonetic G2P. While MeloTTS and SNAP C++ support multilingual processing (Korean, Japanese, English), this repository provides an end-to-end integrated module optimized for high-fidelity speech synthesis.

📄 Technical Whitepaper: 🇬🇧 English Version | 🇰🇷 Korean Version
🤗 Live Demo: Hugging Face Spaces

---

1. Quick Start: Installation & Demo Guide

Linux

# 1. Clone the repository
git clone https://github.com/snap-libs/snap_voice.git
cd snap_voice

# 2. Install Python dependencies
pip install -r requirements.txt

# 3. Setup C++ shared library and model assets
#    (Sparse checkout lib & scripts from snap_cpp, run setup script for Korean model)
git clone --depth 1 --filter=blob:none --sparse https://github.com/snap-libs/snap_cpp.git ../snap_cpp
git -C ../snap_cpp sparse-checkout set lib/linux/x64/v1.0.0 scripts
mkdir -p bin && cp -r ../snap_cpp/lib/linux/x64/v1.0.0/* ./bin/
../snap_cpp/scripts/snap_init.sh -y --lang ko

# 4. Run speech synthesis inference demo
python infer_demo.py --text "여기서 3번 버스를 타고 3번 갈아타야 합니다." --output output_kr.wav

Windows (PowerShell)

# 1. Clone the repository
git clone https://github.com/snap-libs/snap_voice.git
cd snap_voice

# 2. Install Python dependencies
pip install -r requirements.txt

# 3. Setup C++ shared library and model assets
#    (Sparse checkout lib & scripts from snap_cpp, run setup script for Korean model)
git clone --depth 1 --filter=blob:none --sparse https://github.com/snap-libs/snap_cpp.git ../snap_cpp
git -C ../snap_cpp sparse-checkout set lib/windows/x64/v1.0.0 scripts
New-Item -ItemType Directory -Force bin
Copy-Item ../snap_cpp/lib/windows/x64/v1.0.0/* bin/ -Recurse -Force
powershell -ExecutionPolicy Bypass -File ..\snap_cpp\scripts\snap_init.ps1 -Yes -Lang ko

# 4. Run speech synthesis inference demo
python infer_demo.py --text "여기서 3번 버스를 타고 3번 갈아타야 합니다." --output output_kr.wav

---

2. Technical Architecture & Key Features

A. Direct BERT Hidden State Export

• Consumes raw 768-dimensional BERT hidden state tensors ([seq_len, 768]) directly from the SNAP C++ SDK's ONNX session via C-API (snap_get_bert_features), eliminating separate PyTorch BERT model instantiation in Python and saving memory.

B. Context-Aware Text Normalization & G2P

• Delegates text normalization, numeral disambiguation (Sino-Korean vs Native-Korean numerals), and Korean phonetic G2P rules to the high-performance SNAP C++ Engine (snap_process).

C. In-Memory Option Overriding

• Supports per-sentence dynamic configuration overriding using in-memory JSON C-API (snap_process_ext) without disk I/O operations.

---

3. Directory Layout

snap_voice/
 ├── melo/                         # MeloTTS backend core package
 ├── docs/                         # Technical whitepapers (English & Korean)
 ├── snap_wrapper.py               # SNAP C++ SDK ctypes C-API bindings
 ├── infer_demo.py                 # End-to-end CLI inference demo script
 ├── requirements.txt              # Pure Python dependencies
 └── README.md                     # Main documentation

---

4. Python API Usage

End-to-End Speech Synthesis

from melo.api import TTS

# 1. Initialize MeloTTS with integrated SNAP C++ Native Pipeline
#    (Automatically binds SNAP C++ ITN, WordPiece G2P, and INT8 ONNX BERT)
model = TTS(language="KR", device="auto")

# 2. Synthesize speech directly to WAV file
text = "여기서 3번 버스를 타고 3번 갈아타야 합니다."
model.tts_to_file(
    text=text,
    speaker_id=model.hps.data.spk2id["KR"],
    output_path="output_demo.wav"
)

(Optional) Direct SNAP C++ SDK Inspection

from snap_wrapper import SNAPEngineManager

# Direct access to SNAP C++ ITN and raw 768-dim BERT hidden state tensor
manager = SNAPEngineManager()
engine = manager.get_engine(lang="ko")

normalized_text = engine.process("여기서 3번 버스를 타고 3번 갈아타야 합니다.")
bert_tensor, word2ph = engine.get_bert_features("여기서 3번 버스를 타고 3번 갈아타야 합니다.")
print(f"Normalized: {normalized_text}")
print(f"BERT Tensor Shape: {bert_tensor.shape}")  # [1, 768, seq_len]

---

5. Related Links & Repositories

• 🏠 SNAP Project Homepage : Official SNAP project website & portal
• 🌐 snap-libs/snap : SNAP main hub repository
• ⚡ snap-libs/snap_cpp : High-performance C++ ITN / G2P / BERT Hidden State SDK
• 🎙️ snap-libs/snap_voice : Multilingual end-to-end speech synthesis modules

---

6. Community & Contact

We welcome questions, feedback, bug reports, and collaboration inquiries!

• ✉️ Official Email: snap.leejh@gmail.com
• 🐛 Bug Reports & Issues: https://github.com/snap-libs/snap_voice/issues
• 💬 GitHub Discussions: https://github.com/snap-libs/snap_voice/discussions
• 🤗 Hugging Face Spaces: https://huggingface.co/spaces/softguy777/snap_voice_demo