whatmodel.sh / tts / VoxCPM2
VoxCPM2
↗ share
+ contribute
VoxCPM2 is an open-weights text-to-speech model from OpenBMB, released April 2026. Its published size is 2B parameters. Licence: apache-2.0.
Catalogue identity notes
openbmb/VoxCPM2 as named by the evaluating reports. Not VoxCPM (v1/1.5).
Model information sources
Release date ↗ · Parameter count ↗
Maker's model card; the official GitHub README (f0c787f) also says 'a **2B** parameter model' and lists 'Backbone Parameters | 2B'. Rounded maker figure; whether the AudioVAE is included is not stated.
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 VoxCPM2, out of 148 with a score.
| # | model | score | where |
|---|---|---|---|
| 25 |  Qwen3-TTS 12Hz 1.7B Base | 65±1 | local |
| 26 | VoxCPM 0.5B | 64±1 | local |
| 27 | VoxCPM2 | 64±1 | local |
| 28 |  Fun-CosyVoice3-0.5B-2512 | 64±1 | local |
| 29 |  Wayu-Paxa-TTS-Edge | 64±4 | local |
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
voice cloning▁·▁▁▁▁▁▁▇█▃▃ #7 of 37 · 32% of its score · 9 tests
Beats Gemini 3.1 Flash TTS Preview, LongCat-AudioDiT-1B, OmniVoice here, though they rank higher overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Seed-TTS test-en · speaker similarity (WavLM-large SV) lab +1 | 75.3% ✎ | █·█·██·▄·▄·▄ beats 9/10 | 8.1% |
| Seed-TTS test-zh · speaker similarity (WavLM-large SV) lab | 79.4% ✎ | ▃··▃█▅▅··▃·▃ beats 9/10 | 5.4% |
| Seed-TTS test-zh · speaker similarity (ERes2Net SV) | 75.6% ✎ | █·····██···█ beats 0/3 | 3.4% |
| Seed-TTS test-en · speaker similarity (ERes2Net SV) | 72.5% ✎ | █····█··█··█ beats 0/3 | 3.4% |
| MiniMax multilingual test set · 21-language mean error (UltraEval-Audio common set) lab | 7.1% ✎ | █··█·····█·█ beats 1/3 | 3.1% |
| Seed-TTS test-en WER (transcriber not stated) | 1.5% ✎ | ▃··········█ beats 3/3 | 3.1% |
| Seed-TTS test-en · Whisper large-v3 WER lab | 1.6% ✎ | ▃··▃▅··▅·█▃▅ beats 10/11 | 2.2% |
| Seed-TTS test-zh · Paraformer CER lab | 0.9% ✎ | ▂··▄··▂▂█·▄▂ beats 10/11 | 1.9% |
1 more test
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Seed-TTS test-en · UTMOS (predicted MOS) | 2.42 / 5 ✎ | █·····██···█ beats 0/3 | 1.2% |
other tests14 not in a task · 68% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| CV3-Eval hard-en · WER lab +1 | 7.2% ✎ | █··██·██████ beats 7/8 | 13% |
| CV3-Eval hard-zh · CER lab +1 | 8.3% ✎ | ▃▃·▃··▃▃█··▃ beats 7/8 | 8.8% |
| CV3-Eval hard-en · speaker similarity (ERes2Net SV) lab +1 | 69.1% ✎ | ▄█▄·▄··▄··▄█ beats 4/8 | 6.6% |
| CV3-Eval ZH (zero-shot Chinese) · CER lab | 3.5% ✎ | █▄···█▄·█··▄ beats 5/8 | 5.8% |
| ZTTS1-Eval in-the-wild · 10-language mean WER | 3.6% ✎ | ▄·····█··▄██ beats 2/7 | 5.0% |
| CV3-Eval hard-en · DNSMOS (predicted quality) | 3.73 / 5 ✎ | █·▄·▄▄·▄···▄ beats 1/6 | 4.8% |
| ZTTS1-Eval Clean (FLEURS-R) · 9-language mean WER | 5.1% ✎ | ▄·····▄█▄█·▄ beats 1/7 | 4.4% |
| ZTTS1-Eval in-the-wild · 10-language mean MSR-UTMOS | 2.40 / 5 ✎ | █··▄█▄····▄▄ beats 0/7 | 4.3% |
6 more tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| CV3-Eval multilingual voice clone · 9-language mean error | 6.2% ✎ | ▄···▄··▄··██ beats 2/6 | 4.0% |
| ZTTS1-Eval Clean (FLEURS-R) · 9-language mean MSR-UTMOS | 3.07 / 5 ✎ | ▃▃█···▃···▃▃ beats 0/7 | 3.7% |
| CV3-Eval protocol, 7 added languages · mean WER (Qwen-Audio 3.0 extension) | 4.2% ✎ | █··█······██ beats 1/3 | 2.8% |
| CV3-Eval EN (zero-shot English) · WER lab | 5.3% ✎ | ▃·······▃▅█▅ beats 5/8 | 1.9% |
| ZTTS1-Eval · FLEURS-R English TTSDS2 prosody | 88.6% ✎ | █··███·····█ beats 4/4 | 1.7% |
| Arabic TTS Arena · community blind A/B rating, 2026-09-24 snapshot | 1064.0 ✎ | ▂▂··▃·█▅···▂ beats 10/12 | 0.8% |
$ where to run it
Other ways to run it
version
- Run it as published · needs ~2.3 GB, Mac only · from github.com/Blaizzy
pip install mlx-audio && python -m mlx_audio.tts.generate --model mlx-community/VoxCPM2-4bit --text "Hello world" --play
Official weights and documentation: huggingface.co/openbmb ↗
$ how the score was worked out · UltraEval-Audio (OpenBMB) · T… 42%, Qwen-Audio-3.0-TTS technical… 31%, ZONOS2 Technical Report (ZTTS… 17% and 3 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 |
|---|---|---|
 UltraEval-Audio (OpenBMB) · TTS replication results · lab-reported | 10 | 42% |
 Qwen-Audio-3.0-TTS technical report | 8 | 31% |
| ZONOS2 Technical Report (ZTTS1-Eval results) | 4 | 17% |
 nineninesix/gepard-1.0 model card (Seed-TTS-eval benchmark table) | 3 | 7% |
| ZTTS1-Eval results snapshot (Zyphra) | 1 | 2% |
| Arabic TTS Arena (Navid AI) · leaderboard snapshot 2026-09-24 | 1 | 1% |
| prior · pull toward the middle | 0% | |
| weighted average | = 64.0 |
$ 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