whatmodel.sh / tts / Tacotron 2 + HiFi-GAN (VCTK, VITS study)
Tacotron 2 + HiFi-GAN (VCTK, VITS study)
↗ share
+ contribute
Tacotron 2 + HiFi-GAN (VCTK, VITS study) is a closed-weights text-to-speech model from VITS study (public implementations), released June 2021.
Catalogue identity notes
Comparison-only. Exact VCTK multi-speaker system in VITS (ICML 2021) Table 3 row 'Tacotron 2 + HiFi-GAN': NVIDIA Tacotron 2 implementation with broadcast speaker embedding concatenated to the encoder output, trained on VCTK (43,470 training clips, 22 kHz) by the VITS authors (§3.1, §4.2), HiFi-GAN second stage without fine-tuning. Study-specific weights, not released; distinct from tacotron2-hifigan-ljspeech. Needed so vits2021-vctk-mos (today a run of one, vits-vctk) has 3 models; the fine-tuned-HiFi-GAN rows (3.19, 3.82) are left out as unnecessary.
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 Tacotron 2 + HiFi-GAN (VCTK, VITS study), out of 148 with a score, then more from VITS study (public implementations).
| # | model | score | where | ||
|---|---|---|---|---|---|
| 140 |  SpeechT5 TTS | 23±3 | local | ||
| 141 |  MetaVoice 1B v0.1 | 20±3 | local | ||
| 142 | Tacotron 2 + HiFi-GAN (VCTK, VITS study) | 19±6 | comparison only | ||
| 143 |  YourTTS | 18±3 | local | ||
| 144 |  MMS TTS Turkish | 18±4 | local | ||
| more from VITS study (public implementations) | |||||
| 135 | Glow-TTS + HiFi-GAN (VCTK, VITS study) | 28±6 | comparison only | ||
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
other tests1 not in a task · 88% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| VITS 2021 VCTK human MOS | 3.14 / 5 ✎ | █····█·····█ beats 0/2 | 88% |
$ where to run it
No verified API or download recorded yet · add one
Weights and documentation: github.com/NVIDIA ↗
$ how the score was worked out · VITS: ICML 2021 original list… 88%
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 |
|---|---|---|
 VITS: ICML 2021 original listening studies | 1 | 88% |
| prior · pull toward the middle | 12% | |
| weighted average | = 19.4 |
$ 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