whatmodel.sh / tts / Vevo (Amphion)
Vevo (Amphion)
↗ share
+ contribute
Vevo (Amphion) is an open-weights text-to-speech model from Amphion (CUHK-Shenzhen), released December 2024. Its published size is 1.32B parameters. Licence: cc-by-nc-4.0.
Catalogue identity notes
amphion/Vevo (vq8192 tokenizer, PhoneToVq8192 AR model, Vq8192ToMels flow matching, vocoder), the ICLR 2025 Vevo 1.0 release; distinct from Vevo1.5 and Vevo2.
Model information sources
OUR SUM of the maker-stated components on the official card for the Vevo-TTS path: PhoneToVq8192 AR transformer (740M) + Vq8192ToMels flow-matching transformer (330M) + Vocos-based vocoder (250M) = 1.32B. The vq8192 content-style tokenizer is not sized and is excluded. If a single stated figure is preferred, the AR model alone is 0.74.
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 Vevo (Amphion), out of 148 with a score, then more from Amphion (CUHK-Shenzhen).
| # | model | score | where | ||
|---|---|---|---|---|---|
| 95 | Amphion VALL-E (LibriLight 6k) | 47±14 | local | ||
| 96 |  Piper (en_US lessac medium) | 47±5 | local | ||
| 97 | Vevo (Amphion) | 46±3 | local | ||
| 98 |  Qwen3-TTS 12Hz 0.6B CustomVoice | 46±6 | comparison only | ||
| 99 |  Dia2 2B | 45±6 | local | ||
| more from Amphion (CUHK-Shenzhen) | |||||
| 53 | MaskGCT | 59±1 | local | ||
| 121 | Amphion VALL-E v2 recipe (MLS) | 36±3 | local | ||
| 146 | Amphion NaturalSpeech 2 (LibriTTS) | 14±3 | local | ||
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
other tests9 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) | 4.01 / 5 ✎ | ▅·█▅▅·█▃▃▅·▅ beats 14/17 | 61% |
| SAP voice reconstruction · human BWS reconstruction preference | 836.0 ✎ | █▄█····▄··▄▄ beats 2/7 | 11% |
| TTSDS2 listening test · naturalness MOS (4 English domains) | 3.36 / 5 ✎ | ▅··▃▅··▅█▅██ beats 16/17 | 6.3% |
| TTSDS2 multilingual benchmark · Chinese overall (YouTube, 2025 round) | 86.8% ✎ | ▄▄·····▄··▄█ beats 2/5 | 6.1% |
| TTSDS2 multilingual benchmark · English overall (YouTube, 2025 round) | 86.5% ✎ | ▂▂█·▂·█▆▂·▄▂ beats 10/17 | 4.1% |
| CodecMOS-Accent VCTK voice clone · human speaker-similarity MOS (S-SPK-SIM) | 4.48 / 5 ✎ | ▄▄█·▄▄▄▄▄·▄█ beats 9/11 | 2.6% |
| CodecMOS-Accent VCTK voice clone · human naturalness MOS (S-NAT) | 3.66 / 5 ✎ | █··▄▄██▄·▄▄▄ beats 3/11 | 2.5% |
| SAP voice reconstruction · TTSDS dual-reference distributional score | 84.2% ✎ | ▄····▄▄▄▄▄·█ beats 4/7 | 1.9% |
1 more test
| test | result | vs the others in its run | counts |
|---|---|---|---|
| SAP voice reconstruction · WER of reconstructed speech (Whisper turbo) | 34.8% ✎ | ▄▄▄··█▄··▄·▄ beats 1/7 | 0.2% |
$ where to run it
Other ways to run it
- Run it as published · needs ~6.78 GB · from github.com/open-mmlab
git clone https://github.com/open-mmlab/Amphion.git && cd Amphion && conda create -n vevo python=3.10 && conda activate vevo && pip install -r models/vc/vevo/requirements.txt && python -m models.vc.vevo.infer_vevotts
Official weights and documentation: huggingface.co/amphion ↗
$ how the score was worked out · TTSDS2 listening test results… 68%, An Evaluation Framework for T… 13%, TTSDS2 multilingual benchmark… 10% 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 | 68% |
 An Evaluation Framework for TTS Voice Reconstruction (SAP dysarthric speakers; arXiv:2606.21343v1) | 3 | 13% |
| TTSDS2 multilingual benchmark results (ttsds/leaderboard_data) | 2 | 10% |
| CodecMOS-Accent: listening-test MOS of zero-shot TTS across English accents (arXiv:2603.14328v1) | 2 | 5% |
| prior · pull toward the middle | 4% | |
| weighted average | = 46.2 |
$ 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