whatmodel.sh / tts / FireRedTTS (v1)
FireRedTTS (v1)
↗ share
+ contribute
FireRedTTS (v1) is an open-weights text-to-speech model from FireRedTeam (Xiaohongshu), released September 2024. Its published size is 0.4B parameters. Licence: apache-2.0.
Catalogue identity notes
FireRedTeam/FireRedTTS (2024 foundation model). Distinct from FireRedTTS-1S and FireRedTTS-2.
Model information sources
FireRedTTS technical report by the maker (linked from the official FireRedTeam/FireRedTTS card). Covers the text-to-speech language model only; the SAST speech tokenizer and the token-to-waveform module sizes are not stated. The HF repo was later updated (2024-10-17 rich-punctuation model); the report does not say whether that update changed the LM size.
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 FireRedTTS (v1), out of 148 with a score, then more from FireRedTeam (Xiaohongshu).
| # | model | score | where | ||
|---|---|---|---|---|---|
| 88 |  Fish Speech 1.2 | 49±14 | local | ||
| 89 |  SILMA TTS v1 | 48±10 | local | ||
| 90 | FireRedTTS (v1) | 48±3 | local | ||
| 91 |  Vokan | 48±14 | comparison only | ||
| 92 |  Orpheus TTS 3B 0.1 (finetune-prod) | 47±5 | api $15.00/M chars +local | ||
| more from FireRedTeam (Xiaohongshu) | |||||
| 12 | FireRedTTS3 (base) | 67±3 | local | ||
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
voice cloning▁·▁▁▁▃▁▁▇█▃▃ #32 of 37 · 64% of its score · 5 tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Seed-TTS test-en · speaker similarity (WavLM-large SV) | 46.0% ✎ | ▃·▂▂▂█▆▃▅··▂ beats 0/30 | 19% |
| Seed-TTS test-zh · speaker similarity (WavLM-large SV) | 63.0% ✎ | ▃·▂··▂·█▄·▃▄ beats 0/25 | 18% |
| Seed-TTS test-en · human speaker-similarity MOS (SMOS) | 2.36 / 5 ✎ | ▄········█·▄ beats 0/3 | 15% |
| Seed-TTS test-en · Whisper large-v3 WER | 3.8% ✎ | ▁···▃▂▂▂█▆▄▂ beats 1/35 | 7.2% |
| Seed-TTS test-zh · Paraformer CER | 1.5% ✎ | ▁·▂▁▁▁▃▃▅█·▂ beats 8/30 | 4.0% |
other tests4 not in a task · 34% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LibriSpeech-PC test-clean (F5-TTS subset) · speaker similarity (WavLM SV) | 47.0% ✎ | ▄·········█▄ beats 0/3 | 21% |
| LibriSpeech-PC test-clean (F5-TTS 4-10 s subset) · Whisper large-v3 WER | 2.7% ✎ | █·····█·█··█ beats 2/3 | 8.8% |
| CodecMOS-Accent VCTK voice clone · human speaker-similarity MOS (S-SPK-SIM) | 3.42 / 5 ✎ | ▄▄█·▄▄▄▄▄·▄█ beats 1/11 | 1.8% |
| CodecMOS-Accent VCTK voice clone · human naturalness MOS (S-NAT) | 3.61 / 5 ✎ | █··▄▄██▄·▄▄▄ beats 2/11 | 1.7% |
$ where to run it
Other ways to run it
- Run it as published · needs ~3.09 GB, NVIDIA only · from huggingface.co/FireRedTeam (official)
git clone -b main https://github.com/FireRedTeam/FireRedTTS.git && cd FireRedTTS && conda create --name redtts python=3.10 && conda activate redtts && conda install pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 pytorch-cuda=12.1 -c pytorch -c nvidia && pip install -e . && pip install -r requirements.txt && huggingface-cli download FireRedTeam/FireRedTTS --local-dir pretrained_models && python -c "import os; import torchaudio; from fireredtts.fireredtts import FireRedTTS; tts = FireRedTTS(config_path='configs/config_24k.json', pretrained_path='pretrained_models'); rec_wavs = tts.synthesize(prompt_wav='examples/prompt_1.wav', text='小红书，是中国大陆的网络购物和社交平台，成立于二零一三年六月。', lang='zh'); rec_wavs = rec_wavs.detach().cpu(); out_wav_path = os.path.join('./example.wav'); torchaudio.save(out_wav_path, rec_wavs, 24000)"
Official weights and documentation: huggingface.co/FireRedTeam ↗
$ how the score was worked out · F5-TTS: A Fairytaler that Fak… 79%, DiSTAR: Diffusion over a Scal… 15%, CodecMOS-Accent: listening-te… 4%
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 |
|---|---|---|
 F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching (arXiv:2410.06885v1) | 6 | 79% |
| DiSTAR: Diffusion over a Scalable Token Autoregressive Representation for Speech Generation (arXiv:2510.12210v2) | 1 | 15% |
| CodecMOS-Accent: listening-test MOS of zero-shot TTS across English accents (arXiv:2603.14328v1) | 2 | 4% |
| prior · pull toward the middle | 3% | |
| weighted average | = 47.7 |
$ 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