Is there free text-to-speech that works offline?

Yes. Supertonic 3 is open-weight, comes with preset voices and runs offline once installed. The catalog has free voices too, and training your own voice is free if you have an NVIDIA GPU.

Do I need a GPU for text-to-speech?

Not for synthesis: Supertonic 3 and SuperClonica run on the CPU. A GPU makes it faster. Training a new voice does require an NVIDIA GPU.

Can I clone my own voice?

Yes, if it is your voice or you have the speaker's consent. Train a voice from a 5- to 10-second recording or commission an author. You get a voice file close to the timbre and manner of the recording.

Which languages can the voices speak?

The 31 languages of Supertonic 3, including English, Spanish, Portuguese, French, German, Russian, Japanese, Korean, Arabic and Hindi. No Chinese yet; for Chinese, Qwen3-TTS or VoxCPM 2 are better options.

Can I install the SuperClonica engine today?

Yes. The packages are on PyPI: superclonica-cpu for the CPU and superclonica-gpu-cuda for an NVIDIA GPU. The engine page lists the install commands and an example, and catalog voices also work in the standard Supertonic 3 client.

TEXT TO SPEECH

Text-to-speech that runs on your own computer

Updated October 23, 2026

SuperClonica is a catalog of voices for Supertonic 3, an open text-to-speech model that runs on an ordinary CPU, plus the SuperClonica engine, which runs the same voices faster. A voice is a single JSON file. You download it once and turn text into speech on your own machine: no server receives your text, and synthesis keeps working when the internet doesn't.

Below: how local TTS works, how fast it is, what is free, how to get a voice based on a recording, and when a cloud service or a bigger model is the better choice.

Browse the voice catalog How to install SuperClonica

How local text-to-speech works with Supertonic 3

Supertonic 3 is an open-weight model from Supertone with about 99 million parameters. It runs through ONNX Runtime and, according to its authors, needs no GPU. It speaks 31 languages, including English, Spanish, Portuguese, French, German, Russian, Japanese and Korean. Chinese is not supported yet.

A voice file stores only the voice profile. To generate audio you also need an engine (the standard Supertonic 3 client or SuperClonica) and the base model, about 400 MB. Installation needs the internet; after that, synthesis runs locally.

Desktop: Windows, macOS and Linux, on the CPU.
Browser: web apps through ONNX Runtime Web.
Mobile: compatible mobile apps.
Your product: games, bots and services on your own servers.

How fast is TTS on a CPU?

Supertone reports a real-time factor of 0.20 for Supertonic 3 on a 16-thread CPU: one second of audio takes 0.2 seconds. For a desktop 8-core processor like the Ryzen 7 7800X3D we estimate 4–6× faster than real time for Supertonic 3 and 6.4–9.7× for SuperClonica. A 10-second sentence is ready in about 1.0–1.6 seconds.

The engine's gain is measured: on average 61% faster than Supertonic 3 on CPU (Ryzen 7 5800X) and 73% faster on a graphics card (RTX 5070 Ti) with the same settings. Pauses come without hiss, phrase endings are tidier, and long texts are voiced in short parts, so words get repeated or cut off less often. SuperClonica reads Supertonic 3 voice files as they are, with no conversion.

Voices: ready-made, free or cloned from a recording

The catalog holds voices from independent authors: game characters, narrators, conversational voices, each with demos. Authors set prices in US dollars, and some voices are free. A purchased voice is for your own projects; reselling it is not allowed.

For a voice close to a specific recording there are two routes. Both take longer than instant cloning in the cloud, but the result is a file that runs fast anywhere. Use only recordings you have rights to, with the speaker's consent.

Commission it: upload a reference, describe the sound and set a budget. Authors send proposals; the budget stays reserved on your balance until you accept the result.
Train it yourself, free: a clean 5- to 10-second recording with its exact transcript, your own computer with an NVIDIA GPU as a worker, and checkpoints to compare. The JSON downloads free of charge.
Sell what you trained: add a cover, a description and a price. Earnings after the platform fee go to your USD balance.

Local TTS or cloud text-to-speech?

Cloud services such as ElevenLabs generate speech on their servers: a browser app, an API, quick cloning, many languages, nothing to install. In exchange, every text goes over the network, usage is counted against a plan, and nothing works offline.

A local voice file reverses those trade-offs: the text stays on your device, there is no network round trip, and nobody meters your usage. The price is setup time and a smaller model: 31 languages, no Chinese, less expressive control.

Fish Audio S2 Pro, OmniVoice and VoxCPM 2 get equal or higher quality marks in our table and cover more languages, but by our estimates they run 2–10× slower than real time on a desktop CPU.

Who uses local text-to-speech

Typical users:

Game developers generating NPC lines on the player's machine.
Video makers and audiobook narrators voicing long texts on a laptop.
Developers who need speech in a browser or on servers without GPUs.
Teams whose documents must not leave the company network.
Voice authors who train voices to sell or to take commissions.

Account and payments

You sign in with GitHub, no passport or phone number needed. Payments are in cryptocurrency, the balance is kept in US dollars, and we never see bank card details.

SuperClonica compared with other TTS models

Our table rates quality, CPU speed, timbre stability and languages. Ratings and speeds are rough estimates for a desktop CPU, not benchmarks.

Fish Audio S2 Pro: top quality mark, 80+ languages, 4–10× slower than real time on CPU.
Qwen3-TTS: Chinese and 3-second cloning, 1–2× slower than real time (0.6B).
OmniVoice: over 600 languages, 2–5× slower than real time.
VoxCPM 2: 30 languages, 48 kHz audio, 2–4× slower than real time.
F5-TTS: zero-shot cloning in Chinese and English, 5–15× slower than real time.