Piper TTS vs SuperClonica: tiny devices or custom voices
Piper is a fast, local neural text-to-speech engine, known as the voice of Home Assistant and as a speech option in the NVDA screen reader. Its voices are trained with VITS and exported to ONNX, and the original project was optimized for the Raspberry Pi 4. In October 2025 the original repository was archived; development continues in OHF-Voice/piper1-gpl under the Open Home Foundation.
SuperClonica also runs speech locally, with a different focus: Supertonic 3 voices as single JSON files, a catalog of author voices and voices made to order. Piper is not in our comparison table, so this page sticks to the facts from its official repositories and says where each one fits.
How Piper works
Each Piper voice is a separate ONNX model with a JSON config, trained for one language. Piper embeds espeak-ng to turn text into phonemes. The current package installs with pip install piper-tts and comes with a command-line tool, a Python API, an HTTP server and a C/C++ library; GPU acceleration through CUDA is optional.
The voice list covers 39 languages, from Arabic and Catalan to Vietnamese and Welsh, including Russian, German, French, Spanish, Portuguese and Chinese. Japanese and Korean are not on it.
Licenses: the engine and each voice
The current engine, piper1-gpl, is under GPL-3.0; the archived original was MIT. That matters if you embed the library in a closed product. Voices carry their own terms: the voice list asks you to read each voice's MODEL_CARD, warns that some voices have restrictive licenses and says Piper is intended for personal use and speech research.
Supertonic 3 uses OpenRAIL-M, one royalty-free license for the model with a list of prohibited uses. Voices bought on SuperClonica are for your own projects and may not be resold.
A new voice: training for Piper or for SuperClonica
Piper does not clone from a short clip. A new voice means training on a dataset of recordings with transcripts. The official guide recommends fine-tuning from an existing checkpoint; most official voices were trained on 24–48 GB cards, and users report success with 8 GB of VRAM.
On SuperClonica a clean recording of 5 to 10 seconds with its exact transcript is enough to start. Your own NVIDIA GPU works as the trainer, you compare checkpoints, and the finished JSON downloads free. Without a GPU, commission the voice from an author or pick a ready voice with demos from the catalog.
Speed and hardware
Piper was built for small devices, and that is its strongest point. We have not benchmarked it and do not rate it in our table.
Supertonic 3 needs a base model of about 400 MB plus the engine. Its usual targets are desktops, laptops, servers and browsers through ONNX Runtime Web; we have no measurements on a Raspberry Pi. On a desktop 8-core CPU we estimate SuperClonica at 6.4–9.7× faster than real time, and the engine is on average 61% faster than the standard Supertonic 3 client on CPU.
When Piper is the better choice
SuperClonica fits when you want a voice that Piper's list does not have, Japanese or Korean speech, or a voice commissioned for your project. Choose Piper for:
- Raspberry Pi and other small devices, smart home assistants and Home Assistant setups.
- Languages Supertonic 3 does not cover, such as Catalan, Welsh, Farsi, Kazakh, Nepali, Swahili or Chinese.
- A C/C++ library or a ready HTTP server for an existing system.
- Projects that are fine with GPL-3.0 and with the license of the chosen voice.