F5-TTS
A research text-to-speech model with zero-shot voice cloning whose code is MIT but whose pretrained weights are non-commercial, because of the dataset they were trained on.
What it does
F5-TTS is a text-to-speech model from a research paper, described in its README as a "Diffusion Transformer with ConvNeXt V2, faster trained and inference". It clones a voice from a reference recording and its transcript, with no training step. The bundled Gradio web app offers basic generation with chunked inference, multi-style and multi-speaker generation, and a voice chat demo. Training and fine-tuning are documented, with a Gradio app for the fine-tuning as well.
The released checkpoints were trained on the Emilia dataset in Chinese and English. A separate document covers custom inference in other languages. Streaming, model size and sample rate are not stated in the README; the vocoder it names runs at 24 kHz.
What it replaces
F5-TTS replaces ElevenLabs for research and personal projects that need a voice clone, and for teams that intend to train their own weights on data they have the rights to.
It does not replace ElevenLabs for a commercial product as shipped, and this is the reason for the open-weights label. The README says: "The pre-trained models are licensed under the CC-BY-NC license due to the training data Emilia, which is an in-the-wild dataset." The code is MIT. So you may use, change and sell software built on the code, and you may not use the published weights commercially.
How it runs
pip install f5-tts, or clone and install in editable mode, or the Docker image at ghcr.io/swivid/f5-tts. There is a command-line tool and a Gradio web app. The README gives PyTorch install paths for NVIDIA, AMD (ROCm, Linux only), Intel GPUs and Apple Silicon. A Triton and TensorRT-LLM runtime is documented, with the project's own figure of a 0.0394 real-time factor on one L20 GPU. Nobody sells a hosted version; a third-party demo on Hugging Face is linked.
Who it fits
A researcher, or a team that will train its own checkpoint on licensed audio and wants a working, permissively licensed pipeline to do it in. A team that wants to ship cloned voices under a clear commercial licence should use Chatterbox.
Limits the README states
Leaving the reference transcript empty makes the tool transcribe the clip itself, "extra GPU memory usage". Some AMD GPUs fail on ROCm 6.x with an invalid device function error. The maintainers write, of the non-commercial weights, "Sorry for any inconvenience this may cause."
Sources
- F5-TTS README and MIT LICENSE: github.com/SWivid/F5-TTS, read 2026-09-16.
- Model card, cc-by-nc-4.0: huggingface.co/SWivid/F5-TTS, read 2026-09-16.
More text to speech and voice cloning
A multilingual text-to-speech system with voice cloning from a 10 to 30 second clip and inline emotion tags, released with its weights under a research licence that requires a separate agreement for commercial use.
A family of text-to-speech models from Resemble AI under MIT, with zero-shot voice cloning from a reference clip and a multilingual model covering 23 languages.
A fast, local neural text-to-speech engine maintained by the Open Home Foundation, used by Home Assistant and the NVDA screen reader, under GPL-3.0.
A Docker image that serves the 82-million-parameter Kokoro model through an OpenAI-compatible speech endpoint, with streaming, voice mixing and nine languages, all under Apache-2.0.
Added September 16, 2026. Found an error on this page? Write to reveneau@licheo.com and it gets fixed in the next weekly pass.
Running one of these inside your own environment, with your own data and your own security rules, is the kind of work Reveneau does. Read how a forward deployed engagement works.