whatmodel.sh / tts / Amphion NaturalSpeech 2 (LibriTTS)
Amphion NaturalSpeech 2 (LibriTTS)
↗ share
+ contribute
Amphion NaturalSpeech 2 (LibriTTS) is an open-weights text-to-speech model from Amphion (CUHK-Shenzhen), released December 2023. Licence: mit.
Catalogue identity notes
amphion/naturalspeech2_libritts: Amphion's reproduction of NaturalSpeech 2 trained on LibriTTS. Microsoft's original NaturalSpeech 2 was never released.
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 Amphion NaturalSpeech 2 (LibriTTS), out of 148 with a score, then more from Amphion (CUHK-Shenzhen).
| # | model | score | where | ||
|---|---|---|---|---|---|
| 144 |  MMS TTS Turkish | 18±4 | local | ||
| 145 |  Antalia 1 | 17±3 | comparison only | ||
| 146 | Amphion NaturalSpeech 2 (LibriTTS) | 14±3 | local | ||
| 147 |  Glow-TTS Common Voice (Turkish) | 11±4 | local | ||
| 148 |  FreyaTTS-small | 9±4 | comparison only | ||
| more from Amphion (CUHK-Shenzhen) | |||||
| 53 | MaskGCT | 59±1 | local | ||
| 95 | Amphion VALL-E (LibriLight 6k) | 47±14 | local | ||
| 97 | Vevo (Amphion) | 46±3 | local | ||
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
other tests6 not in a task · 96% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| TTSDS2 listening test · speaker-similarity MOS (4 English domains) | 2.06 / 5 ✎ | ▅·█▅▅·█▃▃▅·▅ beats 0/17 | 58% |
| VERSA TTS demo · LibriSpeech test-clean UTMOS | 2.35 / 5 ✎ | ▄·····▄·▄█·▄ beats 0/5 | 24% |
| TTSDS2 multilingual benchmark · English overall (YouTube, 2025 round) | 81.2% ✎ | ▂▃█·▂·█▆▂·▄▂ beats 1/17 | 6.2% |
| TTSDS2 listening test · naturalness MOS (4 English domains) | 2.05 / 5 ✎ | ▅··▃▅··▅█▅██ beats 1/17 | 5.8% |
| TTSDS v1 · default four-factor score | 82.1% ✎ | ▂▂····▃·▃▅▂█ beats 3/14 | 1.0% |
| VERSA TTS demo · LibriSpeech test-clean Whisper WER | 4.6% ✎ | ▄··········█ beats 5/5 | 0.7% |
$ where to run it
Other ways to run it
- Run it as published · needs ~1.66 GB, NVIDIA only · from huggingface.co/amphion (official)
git lfs install && git clone https://huggingface.co/amphion/naturalspeech2_libritts && git clone https://github.com/open-mmlab/Amphion.git && cd Amphion && conda create --name amphion python=3.9.15 && conda activate amphion && sh env.sh && mkdir -p ckpts && ln -s ../../../naturalspeech2_libritts ckpts/tts/ && bash egs/tts/NaturalSpeech2/run_inference.sh --text "Hello, world."
Official weights and documentation: huggingface.co/amphion ↗
$ how the score was worked out · TTSDS2 listening test results… 64%, VERSA toolkit paper: TTS demo… 25%, TTSDS2 multilingual benchmark… 6% and 1 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 |
|---|---|---|
 TTSDS2 listening test results (ttsds/listening_test) | 2 | 64% |
 VERSA toolkit paper: TTS demonstration (arXiv:2412.17667v2) | 2 | 25% |
| TTSDS2 multilingual benchmark results (ttsds/leaderboard_data) | 1 | 6% |
| TTSDS v1 published results | 1 | 1% |
| prior · pull toward the middle | 4% | |
| weighted average | = 14.0 |
$ 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