An alternative to Soprano, Kani TTS and Marvis TTS
A new wave of small open text to speech models aims at real-time speech on modest hardware. Soprano 1.1 has 80 million parameters and is English only; Kani TTS 370M speaks six languages including Chinese; Marvis TTS 250M streams speech in English, French and German and can clone a voice from reference audio.
SuperClonica runs Supertonic 3, about 99 million parameters, in 31 languages on an ordinary processor. This guide compares them with facts from their model cards; we have not benchmarked these models, so the speed figures below are their own claims, not our measurements.
Soprano 1.1
An 80-million-parameter Apache 2.0 model that claims up to 2000 times real time on a GPU and 20 times on a CPU, streams with under 250 ms latency on a CPU and needs under 1 GB of memory. Its card says it is English only, does not clone voices and was trained on about 1,000 hours of audio, so rare words may be mispronounced.
Kani TTS 370M
A 370-million-parameter model at 22 kHz for English, German, Chinese, Korean, Arabic and Spanish, needing about 2 GB of GPU memory. Check its license before commercial use: the card’s metadata names the LFM 1.0 license while its text says Apache 2.0.
Marvis TTS 250M
An Apache 2.0 conversational model with a 250M backbone and a 60M audio decoder, built for real-time streaming on devices such as iPhone and iPad. It is optimized for English, French and German and clones a voice from reference audio.
Where SuperClonica fits
Supertonic 3 speaks 31 languages, and every voice speaks all of them, with about 190 catalog voices from narrators to game and anime characters, bought once; the ten stock voices are free and may be used commercially. We estimate the engine at 6.4 to 9.7 times real time on a desktop 8-core CPU. Install it with python -m pip install superclonica-cpu and run superclonica --text "Hello" --voice F1 --lang en --output hello.mp3; any voice page also plays a line free.
Which to choose
Choose Soprano for the fastest English on tiny hardware, Kani for Chinese or Korean with a GPU, Marvis for streaming and cloning in three European languages. Choose SuperClonica for 31 languages on a CPU, ready character voices, voices trained in its studio, and an API with 3 free minutes a day.