OuteTTS, Llasa, MegaTTS3 and MaskGCT: an honest comparison
OuteTTS, Llasa, MegaTTS3 and MaskGCT are open research models that clone a voice from a short recording, and they often come up together when people look for a self-hosted alternative to cloud voices. Before you build on one, two details matter most: which languages it speaks and what its licence allows.
This guide sums up their model cards and explains when SuperClonica, built on the small Supertonic 3 model, is the simpler choice.
OuteTTS 1.0
OuteTTS 1.0 1B is built on Llama 3.2 1B and speaks 23 languages, 12 of them with strong training, cloning a voice from about 10 seconds of audio; a GGUF version runs with llama.cpp. Its licence combines the Llama 3.2 Community License with CC-BY-NC-SA 4.0 for OuteAI’s additions, so commercial use is restricted.
Llasa
Llasa comes in 1B, 3B and 8B sizes built on Llama models, was trained on 250,000 hours of Chinese and English speech and clones from a 16 kHz prompt. It speaks Chinese and English, and its CC BY-NC 4.0 licence prohibits free commercial use.
MegaTTS3 and MaskGCT
ByteDance’s MegaTTS3 is Apache-2.0 and speaks English and Chinese, but its WaveVAE encoder is not released, so it can only clone from pre-extracted latents, and the card says it is primarily intended for academic purposes; on a CPU it is slow. Amphion’s MaskGCT is a fully non-autoregressive zero-shot model trained on about 100,000 hours of English and Chinese speech, under CC-BY-NC 4.0.
When SuperClonica fits better
If you need commercial use, more languages or a computer without a GPU, SuperClonica is simpler: Supertonic 3 runs on an ordinary processor, every voice speaks 31 languages, the ten stock voices are free under OpenRAIL-M, catalog voices cost $2.99 once, and the studio trains your own voice from a 5 to 10 second recording. Try any voice free on its page with 30 to 100 characters, or install the engine with python -m pip install superclonica-cpu.