whatmodel.sh / tts / XTTS v2
XTTS v2
↗ share
+ contribute
XTTS v2 is an open-weights text-to-speech model from Coqui, released November 2023. Its published size is 0.482B parameters. Licence: coqui-public-model-license.
Model information sources
XTTS paper by Coqui (Casanova et al.), which states "XTTS model and checkpoints are publicly available at ... https://huggingface.co/coqui/XTTS-v2/tree/v2.0.2". Value is OUR SUM of the three maker-stated components (13M + 443M + 26M = 482M). The conditioning encoder (Perceiver) and the H/ASP speaker encoder are not given separate counts. If a single stated figure is preferred, the GPT-2 LM alone is 0.443. The paper links tag v2.0.2; the catalogue main revision is the later v2.0.3-era card, same architecture as far as the card says.
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 XTTS v2, out of 148 with a score.
| # | model | score | where |
|---|---|---|---|
| 129 |  Simple-TTS | 31±5 | comparison only |
| 130 |  StyleTTS 2 (LibriTTS) | 30±2 | local |
| 131 | XTTS v2 | 30±2 | local |
| 132 | Tortoise TTS v2 | 29±2 | local |
| 133 |  Parler-TTS Large v1 | 29±3 | local |
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
voice cloning▁·▁▃▁▁▁▁▇█▃▃ #34 of 37 · 12% of its score · 3 tests
Beats StyleTTS 2 (LibriTTS) here, though they rank higher overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| ZS-TTS-Evaluation · English zero-shot cloning (DAPS speakers, FLORES+) · Whisper-large-v3 CER lab | 0.5% ✎ | █······█··██ beats 3/3 | 9.4% |
| Seed-TTS test-en · Whisper large-v3 WER | 3.2% ✎ | ▁···▂▃▂▂█▆▄▂ beats 4/35 | 1.8% |
| Seed-TTS test-zh · Paraformer CER | 2.9% ✎ | ▁·▃▁▁▁▂▃▅█·▂ beats 2/30 | 1.3% |
other tests25 not in a task · 86% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Freya-TR-Eval Turkish TTS intelligibility (WER, faster-whisper large-v3, 8 kHz band) | 3.7% ✎ | ▄··▄··▄···▂█ beats 4/10 | 19% |
| VoiceCraft zero-shot TTS listening test (80) · naturalness MOS | 3.96 / 5 ✎ | █········█·█ beats 1/2 | 8.6% |
| ClonEval · LibriSpeech test-clean voice-clone speaker similarity (WavLM) | 88.8% ✎ | █·█········█ beats 1/2 | 6.7% |
| LibriSpeech test-clean 4-10 s cross-sentence · HuBERT-large WER | 5.1% ✎ | █···█··█···█ beats 1/3 | 6.1% |
| StyleTTS-ZS LibriSpeech-40 listening test · naturalness MOS | 3.68 / 5 ✎ | █·█········█ beats 1/2 | 5.2% |
| LibriSpeech test-clean 4-10 s cross-sentence · SIM-o (WavLM-TDCNN) | 49.3% ✎ | █··········█ beats 2/3 | 4.4% |
| ZS-TTS-Evaluation · English zero-shot cloning (DAPS, FLORES+) · ECAPA2 speaker similarity lab | 64.2% ✎ | █····▄·····▄ beats 3/3 | 4.3% |
| VoiceCraft zero-shot TTS set (250) · speaker similarity (WavLM-TDCNN) | 47.0% ✎ | █····█·····█ beats 1/2 | 4.2% |
17 more tests
$ where to run it
Other ways to run it
version
- Run it as published · needs ~1.88 GB · from github.com/idiap
uv run --python 3.11 --with coqui-tts --with torch --with torchaudio --with torchcodec tts --model_name tts_models/multilingual/multi-dataset/xtts_v2 --text "It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent." --speaker_idx "Ana Florence" --language_idx en --out_path speech.wav --device cpu
Official weights and documentation: huggingface.co/coqui ↗
$ how the score was worked out · Antalia-2 Mini model card (cl… 19%, VoiceCraft: Zero-Shot Speech… 17%, ZS-TTS-Evaluation (Edresson C… 14% and 10 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
$ 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