whatmodel.sh / tts / Pheme 300M
Pheme 300M
↗ share
+ contribute
Pheme 300M is an open-weights text-to-speech model from PolyAI, released January 2024. Its published size is 0.3B parameters. Licence: cc-by-4.0.
Catalogue identity notes
PolyAI/pheme, which the Pheme README describes as 'the larger version (300M)' (the 100M variant is PolyAI/pheme_small). Repository licence CC-BY-4.0.
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 Pheme 300M, out of 148 with a score.
| # | model | score | where |
|---|---|---|---|
| 115 |  Maya1 | 38±6 | local |
| 116 |  Gemini 3.8 Flash TTS | 37±3 | api price unknown |
| 117 | Pheme 300M | 37±3 | local |
| 118 |  VoiceCraft 830M TTSEnhanced | 37±8 | local |
| 119 | Gemini 3.8 Flash-Lite TTS | 37±3 | api price unknown |
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
other tests4 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) | 3.35 / 5 ✎ | ▅·█▅▅·█▃▃▅·▅ beats 9/17 | 82% |
| TTSDS2 listening test · naturalness MOS (4 English domains) | 2.99 / 5 ✎ | ▅··▃▅··▅█▅██ beats 8/17 | 7.2% |
| TTSDS2 multilingual benchmark · English overall (YouTube, 2025 round) | 87.1% ✎ | ▂▂█·▂·█▆▂·▄▂ beats 12/17 | 5.3% |
| TTSDS v1 · default four-factor score | 84.9% ✎ | ▂▂····▃·▃▅▂█ beats 6/14 | 1.0% |
$ where to run it
Other ways to run it
- Run it as published · needs ~3.6 GB, NVIDIA only · from github.com/PolyAI-LDN
git clone https://github.com/PolyAI-LDN/pheme && cd pheme && conda create --name pheme3 python=3.10 && conda activate pheme3 && pip3 install torch torchvision torchaudio && pip3 install -r requirements.txt --no-deps && mkdir -p ckpt/speechtokenizer && wget -P ckpt/speechtokenizer "https://huggingface.co/fnlp/SpeechTokenizer/resolve/main/speechtokenizer_hubert_avg/SpeechTokenizer.pt" "https://huggingface.co/fnlp/SpeechTokenizer/resolve/main/speechtokenizer_hubert_avg/config.json" && wget -P ckpt "https://huggingface.co/fnlp/USLM/resolve/main/USLM_libritts/unique_text_tokens.k2symbols" && export HUGGING_FACE_HUB_TOKEN=YOUR_PRIVATE_TOKEN && git clone https://huggingface.co/PolyAI/pheme ckpt/pheme && mkdir -p ckpt/t2s ckpt/s2a && mv ckpt/pheme/config_t2s.json ckpt/t2s/config.json && mv ckpt/pheme/generation_config.json ckpt/t2s/generation_config.json && mv ckpt/pheme/t2s.bin ckpt/t2s/pytorch_model.bin && mv ckpt/pheme/config_s2a.json ckpt/s2a/config.json && mv ckpt/pheme/s2a.ckpt ckpt/s2a/s2a.ckpt && python transformer_infer.py
Official weights and documentation: huggingface.co/PolyAI ↗
$ how the score was worked out · TTSDS2 listening test results… 89%, TTSDS2 multilingual benchmark… 5%, TTSDS v1 published results 1%
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 | 89% |
| TTSDS2 multilingual benchmark results (ttsds/leaderboard_data) | 1 | 5% |
| TTSDS v1 published results | 1 | 1% |
| prior · pull toward the middle | 4% | |
| weighted average | = 37.4 |
$ 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