whatmodel.sh / tts / StyleTTS 2 (LibriTTS)

StyleTTS 2 (LibriTTS)

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

StyleTTS 2 (LibriTTS) is an open-weights text-to-speech model from Yinghao Aaron Li et al. (Columbia University), released November 2023. Licence: styletts2-pretrained-model-terms.

Catalogue identity notes

yl4579/StyleTTS2-LibriTTS, the multi-speaker zero-shot checkpoint (the LJSpeech checkpoint is a separate single-speaker model). Code MIT; pre-trained models carry a use condition (listeners must be told samples are synthetic unless the voice owner consents).

30.3±2.1#104 of 118 text-to-speech modelsfrom 16 benchmark results across 8 sources

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 StyleTTS 2 (LibriTTS), out of 148 with a score.

#modelscorewhereRAMreleased
128OuteTTS 0.2 500M OuteAI31±6local~0.9GB2024-11
129Simple-TTS ASAPP Research (Justin Lovelace et al.)31±5comparison only2023-10
130StyleTTS 2 (LibriTTS) Yinghao Aaron Li et al. (Columbia University)30±2local~0.9GB2023-11
131XTTS v2 Coqui30±2local~1.9GB2023-11
132Tortoise TTS v2 Tortoise29±2local~4.5GB2022-04

see all text-to-speech models · find the ones that fit your machine

$ what it's good at

voice cloning▁·▁▃▁▁▁▁▇█▃▃ #35 of 37 * · 13% of its score · 1 test

Trails XTTS v2 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
ZS-TTS-Evaluation · English zero-shot cloning (DAPS speakers, FLORES+) · Whisper-large-v3 CER0.7% ✎█······█··██ beats 2/3avg 0.0%▇▇▇▇▇ 13%typical
other tests15 not in a task · 85% of its score
testresultvs the others in its rundifficultycountswhy
TTSDS2 listening test · speaker-similarity MOS (4 English domains)2.93 / 5 ✎▅·█▅▅·█▃▃▅·▅ beats 8/17easy · avg 4.03 / 5▇▇▇▇▇▇▇▇▇▇▇▇▇ 33%at its level
LibriSpeech test-clean 4-10 s cross-sentence · HuBERT-large WER2.5% ✎█···█··█···█ beats 3/3avg 0.7%▇▇▇ 8.4%typical
StyleTTS-ZS LibriSpeech-40 listening test · naturalness MOS4.23 / 5 ✎█·█········█ beats 2/2very easy · avg 4.64 / 5▇▇▇ 7.2%at its level · small run
LibriSpeech test-clean 4-10 s cross-sentence · SIM-o (WavLM-TDCNN)37.5% ✎█··········█ beats 0/3easy · avg 72%▇▇ 6.1%at its level
ZS-TTS-Evaluation · English zero-shot cloning (DAPS, FLORES+) · ECAPA2 speaker similarity47.3% ✎█····▄·····▄ beats 1/3easy · avg 80%▇▇ 5.8%at its level
RVCBench LibriTTS clean voice cloning · ECAPA speaker similarity23.0% ✎▃·····▂▂▂▃█▆ beats 0/15medium · avg 46%▇▇ 5.1%blunt · at its level
StyleTTS-ZS LibriSpeech-40 · HuBERT-large WER1.6% ✎██·········█ beats 2/2avg 0.6%▇▇ 4.9%noisy · small run
SAP voice reconstruction · human BWS reconstruction preference863.1 ✎█▄█····▄··▄▄ beats 4/7avg 995▇▇ 4.8%typical
7 more tests
testresultvs the others in its rundifficultycountswhy
StyleTTS-ZS LibriSpeech-40 listening test · speaker-similarity MOS3.42 / 5 ✎█··█·······█ beats 0/2very easy · avg 4.55 / 5▇ 3.6%at its level · small run
TTSDS2 listening test · naturalness MOS (4 English domains)3.01 / 5 ✎▅··▃▅··▅█▅██ beats 9/17medium · avg 3.30 / 5▇ 2.8%blunt · at its level
TTSDS2 multilingual benchmark · English overall (YouTube, 2025 round)84.7% ✎▂▂█·▃·█▆▂·▄▂ beats 6/17very easy · avg 89%▇ 2.2%blunt
SAP voice reconstruction · TTSDS dual-reference distributional score83.0% ✎▄····▄▄▄▄▄·█ beats 1/7very easy · avg 86%▇ 1.0%blunt
TTSDS v1 · default four-factor score88.0% ✎▂▂····▃·▃▅▂█ beats 14/14very easy · avg 86%▇ 0.4%blunt
SAP voice reconstruction · WER of reconstructed speech (Whisper turbo)6.4% ✎▄▄▄··█▄··▄·▄ beats 7/7avg 17%▇ 0.1%blunt · noisy
RVCBench LibriTTS clean voice cloning · Whisper-medium WER5.0% ✎▂▂▂··▂·▃▃▃█▂ beats 10/15avg 8.6%▇ 0.1%blunt · noisy

$ where to run it

Other ways to run it

  1. Run it as published · needs ~0.91 GB · from github.com/sidharthrajaram
    uv run --python 3.10 --with styletts2 python -c "from styletts2 import tts; my_tts = tts.StyleTTS2(); out = my_tts.inference('Hello there, I am now a python package.', output_wav_file='test.wav')"

Weights and documentation: huggingface.co/yl4579 ↗

$ how the score was worked out · TTSDS2 listening test results… 36%, ZS-TTS-Evaluation (Edresson C… 19%, StyleTTS-ZS: Efficient High-Q… 16% and 5 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

sourceresultsshare of the scorelicence
TTSDS2 listening test results (ttsds/listening_test)2▇▇▇▇▇▇▇▇▇▇▇ 36%MIT
ZS-TTS-Evaluation (Edresson Casanova) · zero-shot TTS metrics of XTTS, YourTTS, TorToiSe and StyleTTS 22▇▇▇▇▇▇ 19%MIT
StyleTTS-ZS: Efficient High-Quality Zero-Shot Text-to-Speech Synthesis with Distilled Time-Varying Style Diffusion (arXiv:2409.10058v1)3▇▇▇▇▇ 16%CC BY-SA 4.0
DiTTo-TTS: Diffusion Transformers for Scalable Text-to-Speech without Domain-Specific Factors (arXiv:2406.11427v2)2▇▇▇▇ 15%CC BY 4.0
An Evaluation Framework for TTS Voice Reconstruction (SAP dysarthric speakers; arXiv:2606.21343v1)3▇▇ 6%CC BY 4.0
RVCBench: Benchmarking Robustness of Voice Cloning Across Modern Audio Generation Models (arXiv:2602.00443v2)2▇▇ 5%CC BY 4.0
TTSDS2 multilingual benchmark results (ttsds/leaderboard_data)1▇ 2%MIT
TTSDS v1 published results1▇ 0%MIT
prior · pull toward the middle▇ 2%
weighted average= 30.3

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

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