VoxCPM2
2B tokenizer-free diffusion-autoregressive TTS with 30 languages, 48 kHz output, voice design and cloning, Apache-2.0 and explicitly commercial-ready.
Overview
Best for: Commercial self-hosted multilingual voices with design and cloning.
At a glance
Only real-time factor published (about 0.3 on RTX 4090), no time to first audio. 48 kHz output.
Text, optional voice description or reference audio
48 kHz
30
Voice design from text; controllable and 'ultimate' cloning
Model card: real-time streaming with RTF ~0.3 on RTX 4090, ~0.13 with Nano-vLLM.
Wherever you deploy it
Your own deployment; no vendor data processing
NVIDIA GPU (RTX 4090 class for the quoted RTF)
Apache-2.0
Features
- streaming
- voice design
- voice cloning
- multilingual without language tags
Pricing
| What | Price | Unit |
|---|---|---|
| Weights | $0 |
Self-hosted: cost is your GPU/CPU time, not per character
Free tier: Open weights
Source: huggingface.co
Setup
- pip install voxcpm soundfile.
- Load VoxCPM.from_pretrained('openbmb/VoxCPM2').
- Call generate.
Endpoint
LocalAuthentication
None
Quick start python
# pip install voxcpm soundfile
from voxcpm import VoxCPM
import soundfile as sf
model = VoxCPM.from_pretrained("openbmb/VoxCPM2", load_denoiser=False)
wav = model.generate(text="VoxCPM2 brings multilingual support and voice design.")
sf.write("output.wav", wav, model.tts_model.sample_rate)Written from the current docs. Check the vendor's SDK version before you ship.
Warnings
RTF is not TTFB
The card quotes real-time factor, not time to first audio; measure first-chunk latency yourself.
Cloning consent
High-fidelity cloning from a short clip needs speaker consent.
2B params
Heavier than Kokoro/Chatterbox-class models; plan GPU capacity per stream.
Plus 3 warnings that apply to all open models APIs. See category warnings.
Limits
None listed.
Models and products
| Name | Status |
|---|---|
| openbmb/VoxCPM2 | Released 2026-04; updated 2026-08 |
Docs and sources
Docs
Sources used
TTFB; generate() argument names beyond the card's example.