whatmodel.sh / tts / E2 TTS Base (F5-TTS reproduction)
E2 TTS Base (F5-TTS reproduction)
↗ share
+ contribute
E2 TTS Base (F5-TTS reproduction) is an open-weights text-to-speech model from SWivid (SJTU X-LANCE), released October 2024. Its published size is 0.3332B parameters. Licence: cc-by-nc-4.0.
Catalogue identity notes
SWivid/E2-TTS E2TTS_Base/model_1200000.safetensors: the F5-TTS authors' reproduction of Microsoft's E2 TTS (the Microsoft model was never released).
Model information sources
F5-TTS paper by the same SWivid team that released SWivid/E2-TTS E2TTS_Base/model_1200000 (paper: base models trained to 1.2M updates). Covers the E2 TTS transformer only; excludes the Vocos vocoder.
Score within the Vapi Humanness · Bradley–Terry rating, 2026-09-25 snapshot comparison set only.
$ how it compares
The text-to-speech models ranked either side of E2 TTS Base (F5-TTS reproduction), out of 148 with a score, then more from SWivid (SJTU X-LANCE).
| # | model | score | where | ||
|---|---|---|---|---|---|
| 48 |  Inworld Realtime TTS-2 | 61±2 | api $25.00/M chars | ||
| 49 |  CosyVoice2-0.5B | 61±1 | local | ||
| 50 | E2 TTS Base (F5-TTS reproduction) | 60±2 | local | ||
| 51 |  MiniMax Speech-02-Turbo | 60±3 | api $60.00/M chars | ||
| 52 |  TTS-1 | 59±4 | api $15.00/M chars | ||
| more from SWivid (SJTU X-LANCE) | |||||
| 40 | F5-TTS v1 Base | 62±3 | local | ||
| 60 | F5-TTS Base (v0) | 57±2 | local | ||
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
voice cloning▁·▁▁▁▁▁▁▇█▃▃ #21 of 37 · 43% of its score · 6 tests
Beats Gemini 3.1 Flash TTS Preview, MOSS-TTS v1.5 (8B), GLM-TTS here, though they rank higher overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Seed-TTS test-en · speaker similarity (WavLM-large SV) lab +1 | 71.0% ✎ | ▁·▂▂▂█▆▃▅··▂ beats 17/30 | 16% |
| Seed-TTS test-en · human speaker-similarity MOS (SMOS) | 3.29 / 5 ✎ | ▄········█·▄ beats 3/3 | 7.9% |
| Seed-TTS test-zh · speaker similarity (WavLM-large SV) lab | 73.0% ✎ | ▁·▂··▃·█▄·▃▄ beats 4/25 | 7.2% |
| Seed-TTS test-en · Whisper large-v3 WER lab +1 | 2.2% ✎ | ▁···▂▂▂▃█▆▄▂ beats 9/35 | 6.8% |
| Seed-TTS test-en · UTMOS (predicted MOS) | 3.20 / 5 ✎ | ▄█▄··▄···▄·▄ beats 4/6 | 3.4% |
| Seed-TTS test-zh · Paraformer CER lab | 2.0% ✎ | ▁·▂▁▁▃▂▃▅█·▂ beats 5/30 | 2.0% |
other tests10 not in a task · 55% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LibriSpeech-PC test-clean (F5-TTS subset) · speaker similarity (WavLM SV) lab +1 | 69.0% ✎ | ▄·········█▄ beats 3/3 | 18% |
| TTSDS2 listening test · speaker-similarity MOS (4 English domains) | 4.37 / 5 ✎ | ▅·█▅▅·█▃▃▅·▅ beats 16/17 | 14% |
| LibriSpeech-PC test-clean (F5-TTS 4-10 s subset) · Whisper large-v3 WER lab +1 | 2.9% ✎ | █·····█·█··█ beats 1/3 | 9.3% |
| LibriSpeech-PC test-clean (F5-TTS subset) · UTMOS (predicted MOS) | 3.47 / 5 ✎ | █···█······█ beats 0/2 | 5.2% |
| SAP voice reconstruction · human BWS reconstruction preference | 1116.7 ✎ | █▄█····▄··▄▄ beats 6/7 | 3.8% |
| TTSDS2 listening test · naturalness MOS (4 English domains) | 3.41 / 5 ✎ | ▅··▃▅··▅█▅██ beats 17/17 | 2.1% |
| TTSDS2 multilingual benchmark · Chinese overall (YouTube, 2025 round) | 89.8% ✎ | ▄▄·····▄··▄█ beats 5/5 | 1.7% |
| TTSDS2 multilingual benchmark · English overall (YouTube, 2025 round) | 89.5% ✎ | ▂▂█·▂·█▆▂·▄▂ beats 15/17 | 1.2% |
2 more tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| SAP voice reconstruction · TTSDS dual-reference distributional score | 86.3% ✎ | ▄····▄▄▄▄▄·█ beats 7/7 | 0.6% |
| SAP voice reconstruction · WER of reconstructed speech (Whisper turbo) | 30.8% ✎ | ▄▄▄··█▄··▄·▄ beats 2/7 | 0.1% |
$ where to run it
Other ways to run it
- Run it as published · needs ~1.33 GB · from github.com/SWivid (official)
pip install f5-tts && f5-tts_infer-cli --model E2TTS_Base --gen_text "Hello, world."
Official weights and documentation: huggingface.co/SWivid ↗
$ how the score was worked out · DiSTAR: Diffusion over a Scal… 40%, F5-TTS: A Fairytaler that Fak… 35%, TTSDS2 listening test results… 16% and 2 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 DiSTAR: Diffusion over a Scalable Token Autoregressive Representation for Speech Generation (arXiv:2510.12210v2) | 7 | 40% |
| F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching (arXiv:2410.06885v1) · lab-reported | 6 | 35% |
 TTSDS2 listening test results (ttsds/listening_test) | 2 | 16% |
| An Evaluation Framework for TTS Voice Reconstruction (SAP dysarthric speakers; arXiv:2606.21343v1) | 3 | 5% |
| TTSDS2 multilingual benchmark results (ttsds/leaderboard_data) | 2 | 3% |
| prior · pull toward the middle | 1% | |
| weighted average | = 59.8 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json