🗣️ Quevedo Voice Model (so-vits-svc-fork)

This repository contains the voice model of the Spanish singer Quevedo, trained for use with the so-vits-svc-fork library (version 3.10.3+ / 4.0.0+).

---

📋 Table of Contents

• Model Specifications
• Repository Structure
• Quick Installation
• CLI Usage
• Python API Usage
• Gradio WebUI Interface
• Hugging Face Spaces Deployment
• Optimization & Tuning Tips
• Ethical Disclaimer

---

📊 Model Specifications

Feature	Value
Speaker ID	quevedo (Index: 0)
Sampling Rate	44100 Hz (44.1 kHz)
Base Architecture	VITS with SoftVC content encoder (HuBERT)
Fork Target Version	so-vits-svc-fork v3.x / v4.x
Pipeline Tag	Audio-to-Audio (Singing/Speech Voice Conversion)

---

📁 Repository Structure

• G_777.pth: Generator model weight file (Git LFS).
• config.json: Model configuration file detailing training hyperparameters and speaker metadata.
• app.py: Sleek, custom-themed interactive graphical interface built with Gradio.
• requirements.txt: Package requirements to run the inference and the Web UI.
• assets/banner.png: Cover image representing the model repository.

---

🛠️ Quick Installation

To run this model on your local machine, set up a Python environment first (Python 3.10 or 3.11 is recommended):

# 1. Clone the repository
git clone https://huggingface.co/lagosproject/quevedo
cd quevedo

# 2. Create and activate a virtual environment
python3 -m venv venv
source venv/bin/activate  # On Windows use: venv\Scripts\activate

# 3. Install dependencies
pip install -r requirements.txt

You must have FFmpeg installed on your system for audio file processing. If you are on Ubuntu/Debian, run sudo apt install ffmpeg. On macOS/Windows, install it via your preferred package manager (e.g. brew install ffmpeg or choco install ffmpeg).

---

💻 CLI Usage

Perform voice conversions directly from your terminal using the svc console script:

# Basic inference
svc infer path/to/input.wav -m G_777.pth -c config.json -s quevedo -o output.wav

# Transposed inference (+3 semitones for high pitch shifts)
svc infer path/to/input.wav -m G_777.pth -c config.json -s quevedo -t 3 -fm crepe -o output.wav

Useful CLI arguments:

• -m / --model-path: Path to the generator checkpoint (G_777.pth).
• -c / --config-path: Path to the configuration file (config.json).
• -s / --spk-list: Speaker name (quevedo).
• -t / --trans: Pitch shift in semitones (negative numbers shift pitch down, positive numbers shift pitch up).
• -fm / --f0-method: Pitch tracking algorithm. Recommended choices: crepe (highest accuracy) or dio (fastest).

---

🐍 Python API Usage

To run voice conversion programmatically inside a custom Python script:

from pathlib import Path
from so_vits_svc_fork.inference.main import infer

# Configure paths
input_audio = Path("vocals_input.wav")
output_audio = Path("quevedo_output.wav")
model_path = Path("G_777.pth")
config_path = Path("config.json")

# Execute inference
infer(
    input_path=input_audio,
    output_path=output_audio,
    model_path=model_path,
    config_path=config_path,
    recursive=False,
    speaker="quevedo",
    transpose=0,              # Adjust if input vocals are in a different octave
    auto_predict_f0=False,    # Keep False for singing (preserves melody), True for speaking
    f0_method="crepe",        # Crepe offers the highest quality pitch extraction
    noise_scale=0.4
)

print(f"Conversion complete: {output_audio}")

---

🎨 Gradio WebUI Interface

The repository contains a sleek, modern, web interface built with Gradio. To run it locally:

python app.py

Once it starts, navigate to http://localhost:7860 in your web browser.

UI Highlights:

• Drag & Drop Upload: Easily upload any WAV/MP3 files or record directly from your microphone.
• Visual Parameters Control: Adjust Pitch Shift, F0 Predictor (crepe, dio, harvest), and Noise Scale interactively.
• Responsive Layout: Designed with a clean glassmorphism dark-mode theme using customized indigo and purple gradients.

---

🚀 Hugging Face Spaces Deployment

To make this model interactive online for public use without requiring local installation:

1. Create a new Space on your Hugging Face account.
2. Select Gradio as the Space SDK.
3. Choose your hardware (a free CPU basic instance is fine, but GPU hardware speeds up inference considerably).
4. Upload all files from this repository to the Space (including app.py, requirements.txt, config.json, G_777.pth and the assets/ folder).
5. The Space will build and deploy the WebUI automatically.

---

💡 Optimization & Tuning Tips

Follow these guidelines to achieve the best output vocal quality for Quevedo:

• Pitch Adjustments: Quevedo has a deep, resonant baritone singing range.
  • If the source vocals are from a female singer, apply a negative pitch shift (typically -8 to -12 semitones).
  • If the source vocals are from a male tenor singer, shift down by -3 to -6 semitones.
  • If the source vocals are already in a deep baritone range, keep the transposition at 0.
• Singing vs. Speech:
  • For songs, disable Auto Predict F0 to maintain the precise pitch notes of the original track.
  • For speech/voice acting, enable Auto Predict F0 so the model generates natural speech intonation.
• Vocal Preparation:
  • Input audio files must be clean, dry acapellas. Background instruments, beats, reverb, noise, or echo will distort the output audio.
  • For long inputs (more than 45 seconds), slice the audio into smaller files to avoid running out of memory (OOM).

---

⚠️ Ethical Disclaimer

This model is intended for artistic, research, and educational purposes. It should not be used to impersonate individuals for fraudulent, misleading, or defamatory purposes.

• If you share covers or musical works created using this model, please label them clearly as AI covers (e.g., "AI Cover").
• Respect local regulations and the moral rights of the original artist. The author of this repository is not responsible for malicious usage by third parties.