Verdict: If your app talks and you are paying a hosted text-to-speech API by the character, VoxCPM2 is the first open-source model that can credibly replace most of that spend. It is a 2-billion-parameter, Apache-2.0 licensed model from OpenBMB that clones voices from a few seconds of audio, invents entirely new voices from a written description, and outputs 48 kHz speech in 30 languages - all on your own GPU. It will not beat ElevenLabs on every sentence, but for the roughly 80% of routine generation most products do, self-hosting now means no per-character bill and no customer text leaving your network.
TL;DR - Last verified: 2026-09-06
- VoxCPM2 (OpenBMB, released April 2026) is a 2B-parameter, Apache-2.0 tokenizer-free TTS model: voice cloning, text-prompted voice design and 30 languages in one checkpoint (OpenBMB/VoxCPM).
- It serves through an OpenAI-compatible
/v1/audio/speechendpoint (vLLM-Omni), so swapping out a hosted API is mostly a base-URL change.- ElevenLabs still wins on convenience and consistent polish; its API quality tier costs $0.10 per 1,000 characters.
- Realistic play: self-host the bulk, keep a hosted API for the hardest 20%.
What is VoxCPM2, exactly?
VoxCPM2 is open-source text-to-speech software from OpenBMB, the team behind the MiniCPM model family. The GitHub repository describes it as a tokenizer-free TTS system: instead of chopping speech into discrete tokens first (which throws away paralinguistic detail), it generates continuous speech representations through an end-to-end diffusion-autoregressive architecture. That design choice is why it handles breathing, pacing and mid-sentence emotion shifts better than token-based models.
The key facts, per the project README and the Hugging Face model card:
- Size: 2 billion parameters (the earlier VoxCPM-0.5B and VoxCPM1.5 checkpoints remain available for lighter hardware).
- License: Apache-2.0, so commercial use is free with no royalties or usage caps.
- Languages: 30, including Arabic, Hindi, Japanese and Turkish, plus nine Chinese dialects including Cantonese and Sichuanese.
- Output: 48 kHz audio.
- Release: April 2026, with a technical report on arXiv.
Three capabilities ship in one checkpoint where you would previously stitch three tools together:
- Plain TTS. Text in, natural speech out, with emotional colouring inferred from the text itself.
- Voice design. Describe a voice in words - "hyped American on caffeine", "calm middle-aged narrator with a slight smile" - and the model invents it. No reference recording needed. If you iterate on personas for a product or game, this turns voice casting from an asset-hunting problem into a configuration value you can throw away in ten seconds.
- Zero-shot voice cloning. Supply a short reference clip of a voice you have permission to use, and the model applies that voice to new text. Supply the reference's transcript as well, and it captures the performance, not just the timbre - pacing and emphasis carry over.
How do you run it locally?
Install is a pip package plus a few gigabytes of weights, per the README and the PyPI package:
- Create a virtual environment with a recent Python and PyTorch build.
-
pip install voxcpm(weights download from Hugging Face on first run). - Generate via the CLI (
voxcpm --text "..." --output out.wav) or the Python API (VoxCPM.from_pretrained("openbmb/VoxCPM2")).
Before you commit hardware, test your actual sentences in the free Hugging Face playground. If the voice fails there, you have saved yourself a deployment project. If you are on Apple Silicon, the community MLX-Audio path runs VoxCPM on Mac, though its documentation notes VoxCPM2 support is not there yet - the 1.x checkpoints are covered.
Can it replace a hosted TTS API in production?
Yes, if you have a GPU box, because the serving story is now boring in the good way. The official-recommended production path is vLLM-Omni: vllm serve openbmb/VoxCPM2 --omni exposes an OpenAI-compatible /v1/audio/speech endpoint with batching and KV-cache management, per the PyPI documentation. That page also reports a real-time factor around 0.13 on an RTX 4090 via the Nano-vLLM-VoxCPM backend, versus roughly 0.3 for the plain PyTorch path - meaning generation runs several times faster than playback.
The practical consequence: if your application already calls OpenAI-shaped audio endpoints, swapping the backend is a base-URL change, not a rewrite. Community wrappers exist too, but the vLLM-Omni route is the one tied to the official project.
On hardware: the model loads in the 8 GB VRAM class, but once you add KV cache, concurrency and longer prompts, a 24 GB card is the comfortable production floor. Budget for that rather than the minimum.
How does it compare with ElevenLabs, Chatterbox and Qwen3-TTS?
Self-hosted VoxCPM2 is not the only credible option. This is the current landscape, with sources per row:
| Option | License / cost | Languages | Voice design | Cloning | Source |
|---|---|---|---|---|---|
| VoxCPM2 (self-hosted) | Apache-2.0, free | 30 + 9 Chinese dialects | Yes, text-prompted | Zero-shot | OpenBMB/VoxCPM |
| ElevenLabs API | $0.10 per 1K chars (v3); $0.05 (Flash/Turbo) | 29-70+ depending on model | No free-form design | Yes, paid tiers | ElevenLabs pricing |
| Chatterbox (Resemble AI) | MIT, free | 23 | Emotion-exaggeration control | Zero-shot, ~5s clip | ResembleAI/chatterbox |
| Qwen3-TTS (Alibaba) | Apache-2.0, free; 1.7B and 0.6B sizes | 10 | Yes, natural-language | Yes, from ~3s | QwenLM/Qwen3-TTS |
The honest trade-offs:
- ElevenLabs is still the quality and convenience benchmark, and its Flash/Turbo tier is down to $0.05 per 1,000 characters. What you pay for is zero ops. What you give up is data control, latency you do not own, and a bill that scales with usage.
- Chatterbox is lighter and simpler, with MIT licensing and strong emotion control, but covers fewer languages and has no free-form voice design.
- Qwen3-TTS is the closest like-for-like rival on voice design and runs in as little as 4 GB VRAM in its 0.6B form, but its language list is a third of VoxCPM2's.
Quality is not uniformly state-of-the-art in every language - expect unevenness in lower-resource languages, and test your specific ones before committing.
What this means for you
If speech is a real product feature - a voice agent, narration pipeline, accessibility layer - the right question is no longer "which hosted API" but "which 80% of traffic can I self-host". VoxCPM2's Apache-2.0 license and OpenAI-compatible serving make it the lowest-friction way to move that share in-house while keeping ElevenLabs or similar as a premium fallback. This fits the same pattern we covered in Voicebox vs ElevenLabs: Free and Local vs Cloud Quality and the broader build-out in Self-Hosted AI Workspaces in 2026. If you want the voice model to sit alongside other local services, see 10 Free Open-Source AI Tools You Can Run Yourself in 2026; Mac users doing dictation specifically should look at our Fluid Voice local dictation guide.
FAQ
Q: Is VoxCPM2 free for commercial use?
A: Yes. Weights and code are released under the Apache-2.0 license, which permits commercial use, modification and self-hosting with no royalties or usage caps (OpenBMB/VoxCPM license).
Q: How much VRAM does VoxCPM2 need?
A: It runs in the 8 GB VRAM class for basic inference, but production use with KV cache and concurrent requests is more comfortable on a 24 GB card. Apple Silicon Macs can run the earlier 1.x checkpoints through the community MLX-Audio project (docs).
Q: Can VoxCPM2 clone any voice?
A: Technically it clones a voice from a short reference clip, and adding the clip's transcript improves how performance and pacing carry over. Legally and ethically, only clone voices you have explicit permission to use.
Q: How many languages does VoxCPM2 support?
A: 30 languages - including English, Arabic, Hindi, Japanese and Turkish - plus nine Chinese dialects such as Cantonese and Sichuanese, per the project README.
Q: Does VoxCPM2 work with OpenAI-compatible clients?
A: Yes. Serving it with vLLM-Omni exposes a drop-in /v1/audio/speech endpoint, so existing OpenAI-style TTS clients work after a base-URL change (PyPI deployment notes).
Q: Is self-hosted TTS cheaper than ElevenLabs?
A: At volume, usually yes. ElevenLabs API quality tiers cost $0.05-$0.10 per 1,000 characters (about a minute of speech), while self-hosted VoxCPM2 costs only GPU time after the fixed hardware spend (ElevenLabs API pricing).
Sources
- OpenBMB/VoxCPM on GitHub - features, license, language list, install
- openbmb/VoxCPM2 on Hugging Face - model weights and card
- VoxCPM2 technical report (arXiv 2606.06928)
- voxcpm on PyPI - serving, vLLM-Omni, performance figures
- ElevenLabs API pricing
- ResembleAI/chatterbox on Hugging Face
- QwenLM/Qwen3-TTS on GitHub
Updates & Corrections
- 2026-09-06 - Initial publication. Verified VoxCPM2 release details, license, language list, serving path and ElevenLabs/Chatterbox/Qwen3-TTS comparison facts against the primary sources linked above.
Researched and drafted with AI agents; reviewed and fact-checked under human editorial oversight. How we work.
Top comments (0)