README.md

metadata

license: mit
title: VoiceCraftr
sdk: gradio
sdk_version: 5.24.0
python_version: '3.10'
emoji: 🎛️
colorFrom: blue
colorTo: green
pinned: true
short_description: Source song plus target voice cover maker.
tags:
  - audio
  - music
  - voice-conversion
  - gradio
  - cover-song

VoiceCraftr

VoiceCraftr is a free Hugging Face version of a Jammable-style cover workflow:

1. Upload a source song.
2. Upload a target voice reference that you own or have permission to use.
3. Generate one cover mix.

This CPU Space uses reference-driven timbre transfer, pitch-range matching, vocal extraction, tuning, compression, and remixing. It does not ship celebrity models or impersonation presets.

Current CPU Build

• Source song upload: MP3, WAV, FLAC, OGG, M4A depending on ffmpeg support
• Target voice reference upload: clean vocal clips work best
• CPU-safe vocal extraction from the source song
• Target-reference spectral envelope transfer
• Optional auto pitch-range matching
• Manual pitch, tuning, reverb, vocal gain, and backing gain controls
• Single final cover output in WAV or MP3
• Consent gate before generation

Why Not Full Neural Cloning On CPU?

Jammable-grade results usually require a neural singing voice conversion backend. The strongest open-source direction is Seed-VC for zero-shot singing voice conversion, or RVC/Applio when using trained licensed voice models. Those stacks are GPU-heavy and fragile on a free CPU Space.

This build prioritizes a working public HF app now, while keeping a clean place to swap in a GPU backend later.

GPU Upgrade Path

For a higher-quality version:

• Upgrade the Space hardware to GPU or ZeroGPU.
• Add a Seed-VC backend for zero-shot source-song plus target-reference conversion.
• Or add a licensed RVC/Applio backend for user-uploaded .pth / .index models.
• Keep the same UI and consent gate.

Responsible Use

Use only songs and target voice samples that you own or have explicit permission to transform. Do not use this Space to impersonate real people, mislead listeners, or publish copyrighted material without permission.