VoxCPM2
OpenBMB's tokenizer-free TTS: a 2B diffusion-autoregressive model covering 30 languages at 48 kHz — clone a voice, or design one from a text description.
About
VoxCPM2 is OpenBMB's flagship open text-to-speech model, and its headline idea is skipping the speech tokenizer entirely: instead of quantizing audio into discrete tokens, it generates continuous speech representations with a diffusion-autoregressive architecture on a MiniCPM-4 backbone. The 2B-parameter model was trained on 2M+ hours of multilingual speech and speaks 30 languages (plus 9 Chinese dialects) with no language tags needed. It offers four generation modes — plain TTS, voice design from a natural-language description ('a young woman, gentle voice'), controllable cloning from a short clip with style guidance, and 'ultimate' cloning from reference audio plus its transcript — and outputs 48 kHz studio-quality audio via its AudioVAE V2, which upsamples from a 16 kHz reference without an external super-resolution step. OpenBMB reports real-time factor around 0.3 on an RTX 4090 (0.13 with Nano-vLLM), and the checkpoint pulled ~755k Hugging Face downloads in a month. Everything is Apache-2.0: pip install voxcpm, weights from Hugging Face or ModelScope, with SFT/LoRA fine-tuning supported.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Provider-reported RTF ~0.3 on RTX 4090 (~0.13 with Nano-vLLM acceleration)
- 754,970 Hugging Face downloads in the prior month (Aug 2026), 100+ public Spaces
- No independent benchmarks found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 9/10
Scored 9/10 (TIER verified, metrics-v1). Evidence: OpenBMB (MiniCPM team) release; HF checkpoint created 2026-04-03 with 1,688 likes and ~755k monthly downloads; 36.7k stars on the project repo; public docs and demo Space. Strengths: Apache-2.0 everything, novel tokenizer-free continuous-space design, 30 languages at 48 kHz, four generation modes including text-description voice design, real-time streaming with SFT/LoRA fine-tuning. Docked for: (1) efficiency and quality claims are provider-reported with no independent benchmarks found; (2) v2-line is young, so long-term maintenance is unproven versus the older 0.5B line. License: Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · OpenBMB VoxCPM2: 2B tokenizer-free diffusion-autoregressive TTS, Apache-2.0
- · HF checkpoint created 2026-04-03: 1,688 likes, ~755k monthly downloads; 100+ public Spaces; project repo 36.7k stars
- · 30 languages + 9 Chinese dialects, no language tags; 48 kHz via AudioVAE V2; 4 modes: TTS, voice design, controllable cloning, ultimate cloning
- · Provider-reported RTF ~0.3 on RTX 4090 (~0.13 Nano-vLLM); pip install voxcpm; SFT/LoRA fine-tuning supported