GUIDE

An alternative to CosyVoice, IndexTTS and Spark-TTS

CosyVoice, IndexTTS and Spark-TTS are among the strongest open speech models built on language-model ideas: they copy a voice from a single clip, follow instructions about emotion and handle Chinese superbly. They are large models, usually run on NVIDIA graphics cards, and their language lists are short.

The free SuperClonica engine runs the smaller Supertonic 3 model on an ordinary processor in 31 languages, with a catalog of ready voices. This guide compares them honestly and says when one of the three is the better pick.

Browse voices Install the SuperClonica engine

What the three do well

For Chinese, its dialects and instant cloning from a clip, they are excellent tools.

  • Fun-CosyVoice 3.0 covers 9 languages, among them Chinese, English, Japanese, Korean, German, Spanish, French, Italian and Russian, plus more than 18 Chinese dialects, and follows instructions about emotion and speed.
  • IndexTTS-2.5 clones a voice from one reference clip in Chinese, English, Japanese, Spanish and Arabic, with fine emotion control.
  • Spark-TTS speaks Chinese and English and clones voices across the two.
  • Licenses differ: CosyVoice is Apache 2.0, IndexTTS uses the bilibili Model Use License Agreement; check each model before commercial use.

Compare the sound free

On the page of any voice, paste a line you voiced with one of them, 30 to 100 characters, choose the language and press Voice it. In a few seconds you can compare, with no account.

What changes with SuperClonica

The text is read as written, and the same seed gives the same take again, which keeps long projects consistent.

  • 31 languages, including Portuguese, Polish, Dutch, Turkish, Ukrainian, Hindi, Indonesian and Vietnamese; Chinese is not among them.
  • Runs on an ordinary processor: in our tests a short phrase took 0.43 s on four threads of a Ryzen 7 5800X.
  • Ready voices: ten stock voices and more than 170 catalog voices with free demos.
  • One command to install: python -m pip install superclonica-cpu, or a Docker image with an HTTP API.

Copying a voice or training one

The three copy a voice from a clip at every generation. In SuperClonica, a voice is trained once in the studio from a clean 5 to 10 second recording, on a computer with an NVIDIA graphics card, or ordered on the exchange from an author who trains it. After that it runs on any processor and in all 31 languages, with a trace of the accent of the original recording. Clone only voices whose owners agreed.

Which one to choose

Choose CosyVoice, IndexTTS or Spark-TTS for Chinese and its dialects, for instant cloning from a clip and for emotion instructions, when you have a graphics card. Choose SuperClonica for the many languages they do not cover, for ready character and narrator voices and for speech at volume on ordinary servers and laptops.