whatmodel.sh / tts / VoxCPM 0.5B
VoxCPM 0.5B
↗ share
+ contribute
VoxCPM 0.5B is an open-weights text-to-speech model from OpenBMB, released September 2025. Its published size is 0.5B parameters. Licence: apache-2.0.
Catalogue identity notes
openbmb/VoxCPM-0.5B, the September 2025 release. Distinct from VoxCPM1.5 and the existing voxcpm2.
Model information sources
Makers technical report (arXiv 2509.24650v1): "a 0.5B-parameter configuration" of the TSLM (MiniCPM-4-0.5B initialised) + RALM + LocEnc/LocDiT generative model; the official README lists it as Backbone Parameters 0.5B. 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 VoxCPM 0.5B, out of 148 with a score.
| # | model | score | where |
|---|---|---|---|
| 24 |  Raon-OpenTTS-0.3B | 65±2 | comparison only |
| 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 |
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
voice cloning▁·▁▁▁▁▁▁▇█▃▃ #13 of 37 · 26% of its score · 4 tests
Beats Gemini 3.1 Flash TTS Preview here, though they rank higher overall. Trails IndexTTS (v1), VoxCPM2 here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Seed-TTS test-en · speaker similarity (WavLM-large SV) lab +2 | 73.2% ✎ | █·█·██·▄·▄·▄ beats 8/10 | 12% |
| Seed-TTS test-zh · speaker similarity (WavLM-large SV) lab +1 | 77.1% ✎ | ▃··▃█▅▅··▃·▃ beats 7/10 | 6.4% |
| Seed-TTS test-en · Whisper large-v3 WER lab +2 | 1.8% ✎ | ▃··▃▅··▅·█▃▅ beats 8/11 | 5.1% |
| Seed-TTS test-zh · Paraformer CER lab +1 | 1.0% ✎ | ▂··▄··▂▂█·▄▂ beats 5/11 | 2.3% |
other tests13 not in a task · 74% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| CV3-Eval hard-en · WER lab +1 | 7.4% ✎ | █··██·██████ beats 6/8 | 12% |
| CV3-Eval hard-en · speaker similarity (ERes2Net SV) lab +1 | 65.0% ✎ | ▄█▄·▄··▄··▄█ beats 0/8 | 10% |
| CV3-Eval EN (zero-shot English) · WER lab +1 | 6.6% ✎ | ▃·······▃▅█▅ beats 2/8 | 10% |
| CV3-Eval ZH (zero-shot Chinese) · CER lab | 3.4% ✎ | █▄···█▄·█··▄ beats 7/8 | 6.9% |
| CV3-Eval hard-zh · CER lab | 10.0% ✎ | ▃▃·▃··▃▃█··▃ beats 2/8 | 5.6% |
| CV3-Eval hard-en · DNSMOS (predicted quality) | 3.78 / 5 ✎ | ▄·····█·▄▄·█ beats 2/6 | 4.6% |
| Raon-OpenTTS-Eval · overall speaker similarity (WavLM) | 64.2% ✎ | ▄·█·█▄·····▄ beats 4/6 | 4.4% |
| Raon-OpenTTS-Eval · human speaker-similarity MOS (overall) | 3.54 / 5 ✎ | █··██······▄ beats 1/6 | 4.0% |
5 more tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| VoxCPM listening test · Chinese speaker-similarity MOS (S-MOS) lab | 4.11 / 5 ✎ | █·······██·█ beats 3/3 | 3.7% |
| VoxCPM listening test · English speaker-similarity MOS (S-MOS) lab | 4.18 / 5 ✎ | ██········██ beats 3/3 | 3.5% |
| VoxCPM listening test · English naturalness MOS (N-MOS) lab | 4.11 / 5 ✎ | █······█··██ beats 2/3 | 3.3% |
| VoxCPM listening test · Chinese naturalness MOS (N-MOS) lab | 4.10 / 5 ✎ | ██·······█·█ beats 2/3 | 2.9% |
| Raon-OpenTTS-Eval · overall WER (4 acoustic conditions) | 9.5% ✎ | ▄··▄█····▄·█ beats 2/6 | 2.2% |
$ where to run it
Other ways to run it
- Run it as published · needs ~5 GB, NVIDIA only · from github.com/OpenBMB (official)
pip install voxcpm && voxcpm design --text "Hello, world." --hf-model-id openbmb/VoxCPM-0.5B --output out.wav
Official weights and documentation: huggingface.co/openbmb ↗
$ how the score was worked out · UltraEval-Audio (OpenBMB) · T… 46%, Raon-OpenTTS: Open Models and… 40%, VoxCPM: Tokenizer-Free TTS fo… 13%
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 | 9 | 46% |
 Raon-OpenTTS: Open Models and Data for Robust Text-to-Speech (arXiv:2605.20830v2) | 9 | 40% |
| VoxCPM: Tokenizer-Free TTS for Context-Aware Speech Generation and True-to-Life Voice Cloning (arXiv:2509.24650v1) · lab-reported | 4 | 13% |
| prior · pull toward the middle | 1% | |
| weighted average | = 64.1 |
not counted (4)
| VoxCPM: Tokenizer-Free TTS for Context-Aware Speech Generation and True-to-Life Voice Cloning (arXiv:2509.24650v1) | Seed-TTS test-en · Whisper large-v3 WER [correct] | 1.8% | the only model in this source's run of the test, so there is nothing to compare it with there; another run of the test counts |
| VoxCPM: Tokenizer-Free TTS for Context-Aware Speech Generation and True-to-Life Voice Cloning (arXiv:2509.24650v1) | Seed-TTS test-en · speaker similarity (WavLM-large SV) [correct] | 72.9% | the only model in this source's run of the test, so there is nothing to compare it with there; another run of the test counts |
| VoxCPM: Tokenizer-Free TTS for Context-Aware Speech Generation and True-to-Life Voice Cloning (arXiv:2509.24650v1) | Seed-TTS test-zh · Paraformer CER [correct] | 0.9% | the only model in this source's run of the test, so there is nothing to compare it with there; another run of the test counts |
| VoxCPM: Tokenizer-Free TTS for Context-Aware Speech Generation and True-to-Life Voice Cloning (arXiv:2509.24650v1) | Seed-TTS test-zh · speaker similarity (WavLM-large SV) [correct] | 77.2% | the only model in this source's run of the test, so there is nothing to compare it with there; another run of the test counts |
$ 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