F5-TTS vs SuperClonica: zero-shot cloning or fast CPU speech
F5-TTS is an open text-to-speech model built on flow matching with a Diffusion Transformer, first released in October 2024 and followed by an improved v1 base model in March 2025. It clones voices zero-shot: give it a short reference clip and its text, and it speaks in that voice. SuperClonica works differently. A voice is trained once into a Supertonic 3 file, which then runs fast on an ordinary CPU.
If you want an F5-TTS alternative for CPU or for commercial use, here is what changes.
How F5-TTS works
F5-TTS is a Diffusion Transformer with ConvNeXt V2, trained on about 100,000 hours of public multilingual data from the Emilia dataset. The official base model is the Chinese–English one. You pass a reference recording and its transcript; if the transcript is left empty, a speech recognition model fills it in at the cost of extra GPU memory.
The community has fine-tuned F5-TTS for other languages, and the repository links to those models. Our table counts the two languages of the official base model.
F5-TTS speed on CPU and GPU
On a GPU F5-TTS is fast: the repository reports a real-time factor of 0.0402 with TensorRT-LLM and 0.1467 with PyTorch on a single NVIDIA L20 at 16 steps.
On CPU it is the slowest model in our table. The ZipVoice paper measured 37.284 at 32 steps on a single thread of a Xeon Platinum 8457C. Assuming a 2.5–7.5× effective gain from a modern 16-thread desktop CPU, we get 5–15× slower than real time: 50–150 seconds for a 10-second phrase. SuperClonica needs about 1.0–1.6 seconds.
Quality and timbre stability
We rate F5-TTS 3 out of 5 for quality and 3 for timbre stability, against 4.5 and 5 for SuperClonica. Zero-shot cloning rebuilds the voice from the reference on every request, so the result depends on the clip you pass. A trained voice file gives you the same starting timbre every time.
License: check before commercial use
The F5-TTS code is under MIT. The pre-trained models are licensed CC-BY-NC because of the Emilia training data, so the official weights cannot be used commercially under the default terms.
Supertonic 3 uses OpenRAIL-M, a royalty-free license with a list of prohibited uses rather than a non-commercial clause. Voices bought on SuperClonica are for your own projects and may not be resold.
When to choose F5-TTS
Choose SuperClonica for products in the 31 Supertonic 3 languages, for CPU-only machines and for long texts where speed adds up. F5-TTS fits these cases:
- Research and non-commercial experiments with zero-shot cloning.
- Chinese or English speech on a GPU, or a community fine-tune for your language.
- A clone from a clip right now, with no training step.