Chatterbox TTS vs SuperClonica: instant cloning or a trained voice file
Chatterbox is a family of open text-to-speech models from Resemble AI, released under the MIT license. Every model clones a voice zero-shot from a short reference clip. The family has Multilingual V3 with 500 million parameters and 23 languages, the English Turbo model with 350 million, the English Nano model with 110 million for CPU use, and a set of single-language fine-tunes.
SuperClonica takes the opposite route: a voice is trained once into a Supertonic 3 file and then runs fast on an ordinary CPU. Chatterbox is not in our comparison table, so we give no speed or quality scores of our own for it.
The Chatterbox models
Multilingual V3 is the general model: Arabic, Danish, German, Greek, English, Spanish, Finnish, French, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Dutch, Norwegian, Polish, Portuguese, Russian, Swedish, Swahili, Turkish and Chinese. Separate fine-tunes cover Chinese, Latin American and European Spanish, Brazilian and European Portuguese, and Hindi.
Turbo is aimed at low-latency English voice agents and understands tags such as [laugh] and [cough]. Nano shares its architecture at 110 million parameters, and Resemble AI reports that it runs 3× faster than real time on 8 CPU cores. The package installs with pip install chatterbox-tts and was tested on Python 3.11.
Cloning from a clip or a trained voice
With Chatterbox you pass a reference recording with each request and the model rebuilds the voice from it. That is the quickest way to a new voice, and the result depends on the clip you pass.
SuperClonica trains the voice once. You need a clean recording of 5 to 10 seconds with its exact transcript and an NVIDIA GPU in your own computer; training is free and the JSON downloads free of charge. Without a GPU, commission an author, with the budget reserved until you accept the result, or pick a ready voice with demos from the catalog. Every request then starts from the same file.
Chatterbox speed on CPU
We have not benchmarked Chatterbox. Resemble AI's figure for Nano is 3× faster than real time on 8 CPU cores; the examples for the larger models run on CUDA.
For SuperClonica we estimate 6.4–9.7× faster than real time on a desktop 8-core CPU. The two figures come from different tests and texts, so treat them as a starting point and measure on your own machine.
Licenses and watermarks
Chatterbox code and weights are MIT, one of the most permissive licenses there is. Every file Chatterbox generates carries Resemble AI's PerTh watermark, an inaudible mark meant to survive MP3 compression and editing, so the speech can be recognized as generated later.
Supertonic 3 is under OpenRAIL-M, a royalty-free license with a list of prohibited uses. Voices bought on SuperClonica are for your own projects and may not be resold.
When Chatterbox is the better choice
SuperClonica fits when you want a fixed voice file, fast CPU speech in the 31 Supertonic 3 languages, or a voice made for you by an author. Chatterbox is better for:
- Instant cloning from a clip, with no training step.
- Chinese, Hebrew, Malay or Swahili, which Supertonic 3 does not cover.
- Expressive English with laughs, coughs and other non-speech tags.
- Projects that need the MIT license for the model weights.