- Home
- Text-to-speech tools
- VoxCPM
VoxCPM
VoxCPM · Beijing, China
What it is
VoxCPM2 is an open-source, free text-to-speech system from VoxCPM, headquartered in Beijing, China. It generates expressive multilingual speech across 30 languages and lists nine Chinese dialects. The system can design a voice from a natural-language description without reference audio and also supports voice cloning. Output audio is 48kHz, reference audio is accepted at 16kHz, and WAV is listed as an export format. VoxCPM provides streaming generation, with the repository reporting a real-time factor as low as about 0.3 on an NVIDIA RTX 4090. Full and LoRA fine-tuning are supported, with adaptation possible from 5–10 minutes of audio. Documented deployment includes Python, CLI, a local web demo, production serving, and on-device inference. vLLM-Omni exposes an OpenAI-compatible speech endpoint, while Nano-vLLM supports concurrent requests. Code and weights use Apache-2.0 and permit commercial use.
For and against
For
- Free and open-source under Apache-2.0
- Supports 30 languages and nine Chinese dialects
- Offers voice design without reference audio
- Includes full and LoRA fine-tuning
- Provides streaming and OpenAI-compatible serving
Against
- Voice design results can vary between runs
- Cloning results can vary between runs
- Unsupported languages may need testing or fine-tuning
- Requires specific Python, PyTorch, and CUDA versions
The verdict
VoxCPM2 offers multilingual speech generation with voice design, cloning, fine-tuning, and several deployment paths. Users should account for variable results and its documented software requirements.
Who it suits
It suits developers building multilingual speech generation, voice design, cloning, or self-hosted services. Python installation requires versions 3.10 or later but below 3.13, PyTorch 2.5.0 or later, and CUDA 12.0 or later.
Plans and prices
Free & Open Source
No paid tiers or usage limits stated
Read 8 Oct 2026 · voxcpm.com
On the record
- API access
- Yesvoxcpm.com · read 20 Sep 2026
- API integration
- Its vLLM-Omni serving option exposes an OpenAI-compatible /v1/audio/speech endpoint.github.com · read 8 Oct 2026
- Audio output
- The model outputs 48kHz audio and accepts 16kHz reference audio.github.com · read 8 Oct 2026
- Commercial use
- The project states its code and weights use the Apache-2.0 license and are free for commercial use.github.com · read 8 Oct 2026
- Deployment
- The repository documents Python and CLI use, a local web demo, production serving, and on-device inference.github.com · read 8 Oct 2026
- Export formats
- WAVvoxcpm.com · read 20 Sep 2026
- Fine-tuning
- It supports full and LoRA fine-tuning, with adaptation possible from 5–10 minutes of audio.github.com · read 8 Oct 2026
- Integrations
- Documented serving options include vLLM-Omni with an OpenAI-compatible speech endpoint and Nano-vLLM for concurrent requests.github.com · read 8 Oct 2026
- Languages
- VoxCPM2 supports 30 languages and lists nine Chinese dialects.github.com · read 8 Oct 2026
- License and cost
- The model weights and code use the Apache-2.0 license, and the repository says they are free for commercial use.github.com · read 8 Oct 2026
- Limitations
- Voice Design and controllable cloning results can vary between runs, and the project says unsupported languages may require testing or fine-tuning.github.com · read 8 Oct 2026
- Maker
- The site identifies Modelbest Technology Co., Ltd. and offers custom solutions through [email protected].voxcpm.com · read 8 Oct 2026
- Platform support
- The README documents Apple Silicon MPS support and C++ inference on CPU, Metal, CUDA, or Vulkan.github.com · read 8 Oct 2026
- Product
- VoxCPM2 is a tokenizer-free text-to-speech system that generates expressive multilingual speech.github.com · read 8 Oct 2026
- Requirements
- The documented Python installation requires Python 3.10 or later but below 3.13, PyTorch 2.5.0 or later, and CUDA 12.0 or later.github.com · read 8 Oct 2026
- Responsible use
- The project forbids using VoxCPM for impersonation, fraud, or disinformation and recommends labeling AI-generated content.github.com · read 8 Oct 2026
- Streaming
- It provides a streaming generation API, and the repository reports real-time factor as low as about 0.3 on an NVIDIA RTX 4090.github.com · read 8 Oct 2026
- Support
- The project links to Feishu, Discord, and its documentation wiki for community discussion and support.github.com · read 8 Oct 2026
