whatmodel.sh / tts / Parler-TTS Large v1
Parler-TTS Large v1
↗ share
+ contribute
Parler-TTS Large v1 is an open-weights text-to-speech model from Hugging Face (Parler-TTS), released August 2024. Its published size is 2.33B parameters. Licence: apache-2.0.
Catalogue identity notes
parler-tts/parler-tts-large-v1 (description-prompted TTS; 2,333,013,362 parameters per HF safetensors metadata). TTSDS2 ran it through a voice-cloning modification of the codebase.
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 Parler-TTS Large v1, out of 148 with a score, then more from Hugging Face (Parler-TTS).
| # | model | score | where | ||
|---|---|---|---|---|---|
| 131 |  XTTS v2 | 30±2 | local | ||
| 132 | Tortoise TTS v2 | 29±2 | local | ||
| 133 | Parler-TTS Large v1 | 29±3 | local | ||
| 134 |  WhisperSpeech v1.95 Medium (7-language) | 28±3 | local | ||
| 135 |  Glow-TTS + HiFi-GAN (VCTK, VITS study) | 28±6 | comparison only | ||
| more from Hugging Face (Parler-TTS) | |||||
| 87 | Parler-TTS Mini v0.1 | 49±14 | local | ||
| 104 | Parler-TTS Mini v1 | 44±6 | local | ||
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
other tests5 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) | 2.93 / 5 ✎ | ▅·█▅▅·█▃▃▅·▅ beats 7/17 | 70% |
| InstructTTSEval EN · mean instruction-following score (APS, DSD, RP) | 45.7% ✎ | ▄····█··▄··▄ beats 0/4 | 14% |
| TTSDS2 listening test · naturalness MOS (4 English domains) | 2.39 / 5 ✎ | ▅··▃▅··▅█▅██ beats 2/17 | 5.9% |
| TTSDS2 multilingual benchmark · English overall (YouTube, 2025 round) | 79.5% ✎ | ▃▂█·▂·█▆▂·▄▂ beats 0/17 | 4.9% |
| TTSDS v1 · default four-factor score | 87.9% ✎ | ▂▂····▃·▃▅▂█ beats 13/14 | 0.9% |
$ where to run it
Other ways to run it
- Run it as published · needs ~9.33 GB · from github.com/huggingface (official)
pip install git+https://github.com/huggingface/parler-tts.git soundfile && python -c "import torch, soundfile as sf; from parler_tts import ParlerTTSForConditionalGeneration; from transformers import AutoTokenizer; d='cuda:0' if torch.cuda.is_available() else 'cpu'; m=ParlerTTSForConditionalGeneration.from_pretrained('parler-tts/parler-tts-large-v1').to(d); t=AutoTokenizer.from_pretrained('parler-tts/parler-tts-large-v1'); g=m.generate(input_ids=t('A female speaker delivers a slightly expressive speech with a very clear recording.', return_tensors='pt').input_ids.to(d), prompt_input_ids=t('Hello, world!', return_tensors='pt').input_ids.to(d)); sf.write('out.wav', g.cpu().numpy().squeeze(), m.config.sampling_rate)"
Weights and documentation: huggingface.co/parler-tts ↗
$ how the score was worked out · TTSDS2 listening test results… 76%, InstructTTSEval: instruction-… 14%, TTSDS2 multilingual benchmark… 5% 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 | 76% |
 InstructTTSEval: instruction-following TTS benchmark (arXiv:2506.16381v1) | 1 | 14% |
| TTSDS2 multilingual benchmark results (ttsds/leaderboard_data) | 1 | 5% |
| TTSDS v1 published results | 1 | 1% |
| prior · pull toward the middle | 4% | |
| weighted average | = 28.8 |
$ 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