Train your own TTS voice on your NVIDIA GPU
Training a text-to-speech model from scratch takes hours of clean recordings and serious hardware. Most people do not need that: they need one new voice that a good model can speak with. That is what SuperClonica trains: a voice for Supertonic 3, an open multilingual model, from a single clean recording of 5 to 10 seconds.
This guide shows what you need, how to connect your own computer with an NVIDIA graphics card as a worker, how training runs in the studio, and how to use or sell the finished voice. Training on your own hardware is free.
What you train, and why it is fast
The base model already knows how to speak 31 languages. Training finds a voice for it: a small JSON file that describes the timbre and manner of your recording. That is why one short reference is enough, and why the finished voice runs offline on an ordinary processor.
Computer requirements
The worker client runs on Windows; the mode is chosen on the worker card, and in Auto mode the worker waits for free memory and switches to Mini after 15 minutes.
- Windows 10/11 (64-bit), an NVIDIA RTX 20 series card or newer (GTX 16 works too), driver 580 or newer.
- Full quality: a graphics card with 12 GB of memory (16 GB is better) and 32 GB of RAM.
- Mini mode: 8 GB of video memory and 16 GB of RAM; the voice comes out slightly less similar.
- 15 GB of free disk space and the Windows page file turned on.
Step 1: create the voice and its reference
In the studio, create a voice and add a reference: a clean recording of 5 to 10 seconds with its exact text, word for word. Record in a quiet room, one speaker, steady volume, in the manner the voice should have. Use your own voice or one whose speaker has agreed.
Step 2: connect your computer as a worker
Open My workers, add a worker and download SuperClonicaWorker.zip: the client with your worker.json, which holds the platform address and your personal token. Unzip it and run SuperClonicaWorker.exe. On the first start it downloads the rest of what it needs, then updates itself. The worker trains only your voices.
Step 3: train and compare checkpoints
Press Train on the reference, choose the number of steps and start; the page estimates how long 500 steps take from past trainings. As training runs, compare checkpoints by their demos and similarity to the reference, and continue training the most promising one.
Step 4: use or sell the voice
Download the JSON free of charge on the training page and voice any text offline with the free SuperClonica engine: python -m pip install superclonica-cpu, then superclonica --text "Hello!" --voice my-voice.json --lang en --output hello.wav. If you like, publish the voice in the catalog with a cover, a description and a price, or take voice commissions from other users on the exchange.