whatmodel.sh / tts / HierSpeech++ v1.1
HierSpeech++ v1.1
↗ share
+ contribute
HierSpeech++ v1.1 is an open-weights text-to-speech model from Korea University (Sang-Hoon Lee et al.), released November 2023. Its published size is 0.204B parameters. Licence: mit.
Catalogue identity notes
HierSpeech++ Large v1.1 (hierspeechpp_v1.1_ckpt.pth, English/Korean, epoch 200, released 2023-11-20) with the LibriTTS-960 text-to-vec model. Distinct from Large v1 (paper version) and the LibriTTS-460/960 checkpoints.
Model information sources
OUR SUM of the maker-stated sizes of the two checkpoints the README's TTS command loads: the hierarchical speech synthesizer (Params 97M in the README table; the paper says 63M are used at inference and 34M only for training) + the LibriTTS-960 text-to-vec model (TTV, 107M). SpeechSR (0.13M) and the denoiser are not included. Observed 2026-10-07.
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 HierSpeech++ v1.1, out of 148 with a score, then more from Korea University (Sang-Hoon Lee et al.).
| # | model | score | where | ||
|---|---|---|---|---|---|
| 109 | EMA Lightning | 40±3 | local | ||
| 110 |  Antalia-2 Mini | 40±4 | comparison only | ||
| 111 | HierSpeech++ v1.1 | 39±3 | local | ||
| 112 |  Trendyol-TTS | 39±3 | comparison only | ||
| 113 |  Magpie TTS Multilingual 357M | 39±6 | local | ||
| more from Korea University (Sang-Hoon Lee et al.) | |||||
| 136 | HierSpeech++ (LibriTTS-960) | 28±5 | local | ||
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
other tests3 not in a task · 95% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| TTSDS2 listening test · speaker-similarity MOS (4 English domains) | 3.48 / 5 ✎ | ▅·█▅▅·█▃▃▅·▅ beats 11/17 | 83% |
| TTSDS2 listening test · naturalness MOS (4 English domains) | 3.08 / 5 ✎ | ▅··▃▅··▅█▅██ beats 10/17 | 7.5% |
| TTSDS2 multilingual benchmark · English overall (YouTube, 2025 round) | 86.3% ✎ | ▂▂█·▂·█▆▂·▄▂ beats 9/17 | 5.3% |
$ where to run it
Other ways to run it
- Run it as published · needs ~2.1 GB, NVIDIA only · from github.com/sh-lee-prml
git clone https://github.com/sh-lee-prml/HierSpeechpp && cd HierSpeechpp && pip install -r requirements.txt && pip install phonemizer && sudo apt-get install espeak-ng && CUDA_VISIBLE_DEVICES=0 python3 inference.py --ckpt "logs/hierspeechpp_eng_kor/hierspeechpp_v1.1_ckpt.pth" --ckpt_text2w2v "logs/ttv_libritts_v1/ttv_lt960_ckpt.pth" --output_dir "tts_results_eng_kor_v2" --noise_scale_vc "0.333" --noise_scale_ttv "0.333" --denoise_ratio "0"
Weights and documentation: github.com/sh-lee-prml ↗
$ how the score was worked out · TTSDS2 listening test results… 90%, TTSDS2 multilingual benchmark… 5%
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 | 90% |
| TTSDS2 multilingual benchmark results (ttsds/leaderboard_data) | 1 | 5% |
| prior · pull toward the middle | 5% | |
| weighted average | = 39.3 |
$ 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