AI voice model · F5-TTS
F5-TTS: The Flow-Matching Voice Cloning Model, Explained
F5-TTS is an open text to speech model that clones a voice from a short reference clip in English and Chinese, with MIT-licensed code and non-commercial CC-BY-NC weights. If you need commercial voiceovers without a GPU, SupaVocal is a hosted alternative from $9 for the first month.
- Code: MIT, weights: CC-BY-NC
- F5-TTS v1 Base: English and Chinese
- SupaVocal: no GPU, commercial use
What is F5-TTS?
F5-TTS is an open-source, zero-shot text to speech model: give it a few seconds of someone speaking plus the text you want, and F5-TTS answers in that voice with no training run. The name comes from its paper title, "A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching", first posted in October 2024 by researchers from the X-LANCE lab at Shanghai Jiao Tong University and partners.
Under the hood F5-TTS is non-autoregressive. Instead of predicting speech token by token, it pads the text with filler tokens to the length of the audio and denoises the whole clip at once with a Diffusion Transformer trained by flow matching. Because the text is simply padded, it needs no duration model, no phoneme aligner and no separate text encoder; a small ConvNeXt V2 block refines the text so it aligns easily with the speech.
The official F5-TTS checkpoints were trained on about 95,000 hours of English and Chinese speech, and ship with a Python package, a command line tool and a local web app.
F5-TTS specs at a glance
The official F5-TTS release as of November 2026, from its paper, README and model card.
| Made by | Researchers from the X-LANCE lab at Shanghai Jiao Tong University and partners |
|---|---|
| Paper | "F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching", October 2024, revised May 2025 |
| Architecture | Diffusion Transformer with ConvNeXt V2 text refinement, trained with flow matching |
| Size | 335.8M parameters: 22 layers, 16 attention heads, 1,024-wide embeddings |
| Training data | About 95,000 hours of English and Chinese from Emilia |
| Current checkpoint | F5TTS_v1_Base, released March 2025, a 1.35 GB file |
| Audio | 24 kHz output through the Vocos vocoder |
| Length per pass | 30 seconds in total, reference clip included; the CLI and web app chunk longer text |
| Code license | MIT |
| Weights license | CC-BY-NC 4.0, because of the Emilia training data |
| Latest package | f5-tts 1.1.22 on PyPI, July 23, 2026 |
| Python | 3.10 or newer, with PyTorch for your CUDA, ROCm, Intel or Apple GPU |
F5-TTS v1 Base vs the original F5-TTS Base and E2 TTS
The repository ships three checkpoints. The v1 Base is the default; the other two remain for older setups.
| F5-TTS v1 Base | F5-TTS Base (v0) | E2 TTS Base | |
|---|---|---|---|
| Released | March 2025 | October 2024 | October 2024 |
| Network | Diffusion Transformer + ConvNeXt V2 | Diffusion Transformer + ConvNeXt V2 | Flat-UNet Transformer |
| Training updates | 1.25M | 1.2M | 1.2M |
| CLI name | F5TTS_v1_Base | F5TTS_Base | E2TTS_Base |
| BigVGAN vocoder option | No, Vocos only | Yes | No |
| Weights license | CC-BY-NC 4.0 | Same | Same |
The maintainers say only that v1 has better training and inference performance; the README and paper list no separate v1 benchmark.
What F5-TTS voice cloning does well
Why people pick F5-TTS, per its paper and README.
Zero-shot cloning
One short reference clip and its transcript are enough; no fine-tuning run is needed for a new speaker.
Fast for a diffusion model
On an RTX 3090 the paper times 10 seconds of speech at about 1.5 seconds with 16 steps, or 3.1 seconds at the default 32.
Code-switching
The base model mixes English and Chinese inside one sentence, a skill the paper highlights.
Multi-voice stories
A .toml file maps tags such as [main] and [town] to different reference clips, so one script can switch speakers.
Speed control
A speed setting stretches or shortens the output, and fix_duration sets an exact total length, reference clip included.
Fine-tune your own
A training guide and a fine-tuning web app let you adapt F5-TTS to a new language or speaker.
How to install F5-TTS locally
To install F5-TTS, follow the four steps its README gives.
- 1
Create an environment
Make a Python 3.10 or newer environment (the README uses conda with 3.11) and install FFmpeg; without it the output can be pure silence.
- 2
Install PyTorch for your device
Install
torchandtorchaudiobuilt for your NVIDIA CUDA, AMD ROCm (Linux only), Intel XPU or Apple Silicon setup. Newer AMD cards need ROCm 7. - 3
Install the package
Run
pip install f5-ttsfor inference only. To train or fine-tune, clone the repository and runpip install -e .inside it. - 4
Generate speech
Run
f5-tts_infer-gradiofor a local web page, orf5-tts_infer-cli --model F5TTS_v1_Base --ref_audio ref.wav --ref_text "..." --gen_text "...". Weights download from Hugging Face on first use.
F5-TTS install and deployment options
Pick the route by who calls the model.
| Route | Runs on | Best for | Notes |
|---|---|---|---|
| Gradio web app | Local GPU or CPU | Trying voices, multi-speaker scripts | f5-tts_infer-gradio, port 7860; includes a voice chat demo |
f5-tts_infer-cli | Local GPU or CPU | Batch narration, shell scripts | Reads a .toml config; writes a WAV file |
| Python API | Local GPU or CPU | Apps and notebooks | from f5_tts.api import F5TTS, then infer() |
| Docker image | NVIDIA GPU | Servers without a Python install | ghcr.io/swivid/f5-tts:main, Gradio on port 7860 |
| Triton + TensorRT-LLM | NVIDIA GPU | High-throughput serving | RTF 0.0394 and 253 ms average latency at concurrency 2, one L20, 16 steps |
A socket server streams audio in chunks for real-time playback.
F5-TTS voice cloning: reference clip rules
F5-TTS voice cloning depends on the reference clip. The official inference guide says:
- Keep it under 12 seconds. Longer clips are cut to about 12 seconds, which can split a word. Leave about 1 second of silence at the end.
- Give the transcript. Pass the exact words in
ref_text. Leave it empty and a speech recognition model transcribes it, using extra GPU memory. - Mind the 30-second window. One pass covers 30 seconds in total, reference included. The CLI and web app split longer text into chunks and cross-fade them by 0.15 seconds.
- Shape the text. Capital letters such as K.F.C. are read letter by letter, extra spaces or commas add pauses, and digits are read in English unless you write them out in Chinese.
The catch is the license. The F5-TTS weights are CC-BY-NC 4.0, so the safe reading is to keep monetized videos, client work and paid audiobooks off them, even though the code is MIT. Clone only your own voice or one whose owner has agreed.
F5-TTS inference settings that matter
Defaults in the official inference code.
| Setting | Default | What it changes |
|---|---|---|
nfe_step | 32 | Denoising steps; 16 is faster, more steps are more accurate but slower |
cfg_strength | 2.0 | How strongly the output follows the text and reference |
sway_sampling_coef | -1.0 | Sway Sampling, the paper's trick for good results with fewer steps |
speed | 1.0 | Below 1 slows the voice down, above 1 speeds it up |
remove_silence | Off | Trims long pauses from the finished audio |
Each setting is a CLI flag and a .toml key.
F5-TTS limitations to plan for
Check these before building a product on F5-TTS.
Do
- Read the weights license: the code is MIT, the F5-TTS checkpoints are CC-BY-NC
- Use a clean reference clip under 12 seconds with a 1-second tail of silence
- Pass the reference transcript yourself to save GPU memory
- Check each community checkpoint's own license before using it
Don't
- Rely on the official checkpoints for paid or monetized work
- Expect the official checkpoint to speak languages beyond English and Chinese
- Feed a long reference clip and hope the cut lands between words
- Clone a voice without the speaker's consent
F5-TTS vs Coqui XTTS v2 vs SupaVocal
Two self-hosted open models next to a hosted studio.
| F5-TTS | Coqui XTTS v2 | SupaVocal | |
|---|---|---|---|
| What it is | Open model, run it yourself | Open model, run it yourself | Hosted voice studio in the browser |
| Languages | English and Chinese; community checkpoints for others | 17 | 30+ stock voices in English and Chinese |
| Reference clip | Under 12 s, plus a transcript or auto-transcription | About 6 s per its model card | At least 3 s, no transcript needed |
| Output | 24 kHz WAV | 24 kHz | MP3 download |
| Commercial use | Not under the CC-BY-NC weights | No, non-commercial model license | Yes, on every plan |
| Hardware | Your GPU (CPU fallback) | Your GPU or CPU | None |
| Cost | Free, plus your hardware | Free, plus your hardware | From $19 a month, $9 the first month |
More on XTTS v2 in our Coqui TTS guide.
An F5-TTS alternative for commercial voiceovers
F5-TTS is the right choice if you have a GPU, want to self-host or fine-tune, and your project is non-commercial. It costs nothing but hardware time.
If you need voiceovers you can sell and don't want to manage PyTorch, CUDA and model licenses, SupaVocal is a separate, hosted AI voice studio; it does not run F5-TTS or any F5-TTS checkpoint. You paste up to 15,000 characters (5,000 with a cloned voice) into text to speech, pick one of 30+ English and Chinese voices and download an MP3, with commercial use on every plan.
- Voice cloning: upload a clip of at least 3 seconds in Create Voice, confirm you have the speaker's consent, and save up to 10 voices. No transcript to type.
- More tools: a voice changer, speech to text for files up to an hour (25 MB), audio translation and sound effects.
Plans start at $19 a month, $9 for the first month, for 200,000 credits. Stock voices use 1 credit per character; cloned voices use 3. See pricing.
Where to go next
SupaVocal tools and guides for anyone weighing F5-TTS against a hosted studio.
Frequently asked questions
What is F5-TTS?
F5-TTS is an open-source zero-shot text to speech model from researchers at Shanghai Jiao Tong University's X-LANCE lab and partners. It clones a voice from a short reference clip using a Diffusion Transformer trained with flow matching, without a duration model or phoneme aligner. The official checkpoints speak English and Chinese; the code is MIT and the weights are CC-BY-NC.
Is F5-TTS free for commercial use?
Not with the official weights. The F5-TTS code is MIT-licensed, but the pretrained checkpoints, including the v1 Base, are CC-BY-NC 4.0 because they were trained on the Emilia dataset. The safe reading is to keep monetized videos, paid audiobooks and client work off them. Community checkpoints carry their own licenses, so check each one.
What changed in F5-TTS v1?
The v1 Base, released in March 2025, is a retrained checkpoint of the same-size Diffusion Transformer with small config changes, trained for 1.25 million updates on the same 95,000 hours of English and Chinese. The maintainers say it trains and infers better, and it is the default model in the CLI and web app.
What languages does F5-TTS support?
The official F5-TTS checkpoints are trained on English and Chinese and can switch between them within a sentence. Volunteers have shared fine-tuned checkpoints for Arabic, Finnish, French, German, Hindi, Italian, Japanese, Latvian, Russian and Spanish, each with its own license and quality. For other languages you would need to fine-tune the model yourself.
How long should an F5-TTS reference clip be?
Keep the reference clip under 12 seconds and leave about one second of silence at the end. Longer clips are cut to roughly 12 seconds, which can split a word. Pass the clip's exact transcript as ref_text; if you leave it empty, a speech recognition model transcribes it, using extra GPU memory.
Does F5-TTS need a GPU?
In practice, yes. The README installs PyTorch for NVIDIA, AMD, Intel or Apple Silicon GPUs, and the paper's real-time factor of 0.15 (at 16 steps) was measured on an RTX 3090. The code falls back to the CPU when no accelerator is found, but no CPU speed figures are listed in the README or paper. For serving at scale, the project documents a Triton and TensorRT-LLM setup on NVIDIA GPUs.
Is there an official F5-TTS website?
The official project is the arXiv paper by its eight authors, the SWivid/F5-TTS code repository with its demo page, and the SWivid checkpoints on Hugging Face. The README links to no paid F5-TTS website, so treat sites selling F5-TTS plans as independent services.
Does SupaVocal use F5-TTS?
No. SupaVocal is a separate, hosted voice studio and does not run F5-TTS or any F5-TTS checkpoint. It is an alternative for people who want text to speech and voice cloning without a GPU: 30+ English and Chinese voices, cloning from a 3-second clip with no transcript, and commercial use on every plan, from $19 a month with the first month at $9.
No GPU needed: try SupaVocal instead of F5-TTS
Paste a script, pick from 30+ English and Chinese voices or clone your own from a 3-second clip. $9 for your first month, commercial use included.
Start creating voiceovers