Northern Forge Labs · Local · Private · GPU

Sing in your own voice.
Keep it on your machine.

Train a singing voice from guided takes, then convert dry vocals with quality presets and pitch control — all on your NVIDIA GPU. Your clips, models, and outputs never leave this computer.

NVIDIA GPU with CUDA Windows 11 + WSL2 or Linux Local run — no cloud inference

Read the local run guide What you get

Consent-only: use your voice (or voices you have explicit rights to). Not a cloud deepfake service. No macOS or AMD GPU support yet.

What you get

Guided vocal training

Walk through fixed singing scripts with readiness tracking so you know when you have enough clean takes. Import an existing model when you already have one ready for convert-only work.

Convert dry vocals

Upload a dry vocal you own, set f0 / pitch shift and a quality preset (match, smooth, or stable), then queue a conversion. Output stays on disk with generation history.

Nothing leaves your computer

Inference is local. There is no shared singing-conversion service, no upload step, and no account required to run the app on your machine.

History & job queue

Conversions are queued and saved with voice name, quality, f0, and paths so you can replay or re-run later without redoing setup.

Sibling of Voice Clone Studio

Voice Clone Studio is local speech TTS (type text → hear your speaking voice). Voice Sing Studio is local singing voice conversion (dry vocal → your singing timbre). Same Labs product family, different job: speech vs song conversion.

This is not a hosted artist marketplace and does not claim commercial cloud cloning. Model licenses and weights are documented in the product README.

Run locally

On a machine with an NVIDIA GPU, running Linux or Windows 11 with WSL2:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
# Optional: CUDA torch + rvc-python for real GPU convert

cp .env.example .env
# UI smoke without GPU:
export VOICE_SING_ENGINE=fake
export VOICE_SING_SSL=0   # optional; HTTP on loopback

bash run.sh
# or: python app.py

Default bind is http(s)://127.0.0.1:7861. Non-loopback binds require auth. First real RVC convert needs GPU-capable torch and the engine deps described in the README / runbook.