🎶 SVS-WORLD

SVS-WORLD: Singing Voice Synthesis with WORLD Vocoder (Windows-ready) RESEARCH Under Development & AI Based, using QWEN

SVS-WORLD is a minimal yet fully functional Singing Voice Synthesis (SVS) system built from scratch using the WORLD vocoder. Designed for beginners and researchers, it runs entirely on Windows (no Linux/WSL needed) and supports real-world datasets with automatic preprocessing.

Unlike deep learning-based systems (e.g., DiffSinger, NNSVS), this project focuses on classical parametric synthesis, making it easy to understand, debug, and extend.

---

✨ Features

• 🖥️ 100% Windows-compatible (tested on Python 3.8–3.11)
• 📂 Supports HTS-style label files with flexible time units:
  • 10 MHz ticks (e.g., 1824671232)
  • Microseconds, milliseconds, or seconds
• 🧹 Automatic label normalization:
  • pau, sil, # → SP (silence)
  • br, bre, AP → AP (breath)
• 🔊 Real silence modeling – no artificial "TV static" in pauses
• 🖼️ Alignment visualization – see F0 vs. phonemes
• 🖱️ Graphical user interface for:
  • Loading .wav + .lab pairs
  • Batch feature extraction
  • Model building
  • Interactive synthesis (phoneme-by-phoneme)
• 🧪 No GPU required – runs on CPU only

---

🛠️ Tech Stack

• Vocoder: WORLD (via pyworld)
• Audio I/O: librosa, soundfile
• GUI: tkinter (built-in)
• Alignment: Manual (HTS-style labels)
• Language: Python 3

---

🚀 Quick Start (Windows)

1. Install dependencies:

   pip install -r requirements.txt
   

Use cases

Educational projects on speech/singing synthesis Custom voice banks for amateur music production Baseline system for SVS research Lightweight alternative to UTAU/DeepVocal

🙌 Acknowledgements

WORLD Vocoder by Masanori Morise pyworld HTS, Sinsy, and OpenUTAU for inspiration