VoxCPM2: Tokenizer-Free Text-to-Speech with 30-Language Support and Voice Cloning

What Problem VoxCPM2 Addresses and Who It Is For

Most text-to-speech systems convert text into discrete tokens representing phonemes or sub-word units, then decode those tokens into audio. This tokenization step introduces a bottleneck: the quality of the synthesis is limited by how well the token vocabulary captures continuous speech characteristics.

VoxCPM is described in the README as 'a tokenizer-free Text-to-Speech system that directly generates continuous speech representations via an end-to-end diffusion autoregressive architecture, bypassing discrete tokenization.' The practical effect is that the model can generate highly natural, expressive audio without the quantization artifacts that discrete token systems introduce.

The primary audience is developers building multilingual voice applications, content creation tools, accessibility tools, and synthetic voice products. The Apache-2.0 license makes the model available for commercial use without royalty payments. Researchers working on speech synthesis and developers who want to integrate voice generation into Python applications are the other main groups this project targets.

Architecture: Diffusion Autoregressive Design and MiniCPM-4 Backbone

VoxCPM2 is a 2B parameter model trained on over 2 million hours of multilingual speech data. The README states it is built on a MiniCPM-4 backbone. The diffusion autoregressive architecture allows the model to generate speech in an autoregressive manner while using diffusion-based refinement, which the README credits with achieving natural and expressive synthesis.

The model outputs audio at 48kHz. The README notes that it accepts 16kHz reference audio and outputs 48kHz studio-quality audio using an AudioVAE V2 asymmetric encode/decode design with built-in super-resolution. This means there is no need for a separate upsampling step when using lower-quality reference recordings.

Streaming is supported. The README gives a real-time factor of approximately 0.3 on an NVIDIA RTX 4090. With acceleration via Nano-vLLM or vLLM-Omni, the RTF drops to approximately 0.13. An OpenAI-compatible API for VoxCPM2 is available through vLLM-Omni using PagedAttention.

Installing VoxCPM2 and Running Your First Synthesis

Installation requires Python 3.10 to 3.12 (not 3.13 or later), PyTorch 2.5.0 or higher, and CUDA 12.0 or higher. The package is on PyPI:

pip install voxcpm

The model weights are hosted on Hugging Face at openbmb/VoxCPM2. The Python API loads them automatically on first use via VoxCPM.from_pretrained:

from voxcpm import VoxCPM
import soundfile as sf

model = VoxCPM.from_pretrained(
  "openbmb/VoxCPM2",
  load_denoiser=False,
)

wav = model.generate(
    text="VoxCPM2 is the current recommended release for realistic multilingual speech synthesis.",
    cfg_value=2.0,
    inference_timesteps=10,
    seed=42,
)
sf.write("demo.wav", wav, model.tts_model.sample_rate)

The generate method returns a waveform array. The soundfile library writes it to a WAV file. The cfg_value parameter controls classifier-free guidance strength, and inference_timesteps controls the number of diffusion steps. The README also shows how to download weights from ModelScope for users who cannot access Hugging Face.

Voice Design: Creating a Voice from a Text Description

VoxCPM2 supports a mode the README calls Voice Design, in which a new voice is generated from a natural-language description alone, with no reference audio required. The format is straightforward: place the voice description in parentheses at the start of the text parameter:

wav = model.generate(
    text="(A young woman, gentle and sweet voice)Hello, welcome to VoxCPM2!",
    cfg_value=2.0,
    inference_timesteps=10,
    seed=42,
)
sf.write("voice_design.wav", wav, model.tts_model.sample_rate)

The description can include attributes such as gender, age, tone, emotion, and pace. The model interprets the description and synthesizes a voice matching those characteristics without any stored voice profile. This is useful for applications where a specific named voice is not required, such as audiobook narration with varied character voices or accessibility tools that need distinct speaker identities.

Where VoxCPM2 Has Limitations

The CUDA 12 requirement is a hard constraint. CPU inference and older CUDA versions are not supported according to the installation requirements documented in the README. This excludes macOS environments and machines without a compatible NVIDIA GPU, which limits deployment to Linux systems with supported hardware.

The model is 2B parameters. Downloading weights from Hugging Face requires sufficient disk space and bandwidth. The initial load time depends on network conditions and local storage speed. The README does not document a quantized or reduced-size variant for lower-resource environments.

The README includes a Risks and Limitations section in the table of contents, but the README does not document the specific risks in that section. The project is Apache-2.0 licensed, which permits commercial use, but any deployment that reproduces another person's voice without their consent raises legal and ethical questions independent of the software license. The README does not document VoxCPM's behavior when given a reference audio of an identifiable person without their consent.

Supported Languages and the CLI

VoxCPM2 supports 30 languages: Arabic, Burmese, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Khmer, Korean, Lao, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, Tagalog, Thai, Turkish, and Vietnamese. The README also lists nine Chinese dialects including Cantonese, Wu, and Sichuan dialect.

A command-line interface is available as the voxcpm command, registered as a project script in pyproject.toml. The CLI entry point is voxcpm.cli:main. The pyproject.toml also lists a timestamps optional dependency group that adds stable-ts, which the README does not describe in the visible text but the dependency name suggests it supports timestamped output.

An alternative to VoxCPM2 for open-source multilingual TTS is Kokoro or Coqui TTS. Both are permissively licensed but do not offer the same architecture (diffusion autoregressive vs autoregressive or flow-matching). VoxCPM2's differentiator is the tokenizer-free approach and the Voice Design feature, which the alternatives do not replicate with the same natural-language prompt interface.

Maintenance Status and Licensing

The last push to the repository was on 2026-09-02. The most recent release, v2.0.3, was published on 2026-05-11 with changes described as fine-tuning validation, runtime stability, and streaming improvements. VoxCPM2 itself was released in April 2026 as a major update from the earlier VoxCPM1.5.

The repository is licensed under Apache-2.0, which permits commercial use, modification, and redistribution without royalty payments. The model weights on Hugging Face are also listed under Apache-2.0 in the README. This is a permissive choice that makes VoxCPM2 available for production products.

The pyproject.toml includes a dev dependency group with pytest and pre-commit tools. A fine-tuning web UI is available as lora_ft_webui.py at the repository root, and a Gradio demo app is at app.py, allowing local testing without writing Python code directly.

Editorial conclusion

Developers who need multilingual TTS with voice cloning in a single open-source package will find VoxCPM2 a direct option, given its Apache-2.0 license, pip installation, and support for Voice Design and Controllable Cloning. Projects that cannot meet the CUDA 12 requirement or cannot use Python 3.10 through 3.12 will need a different path. Before deploying a cloned voice in a production product, verify that the intended use complies with applicable regulations on synthetic voice reproduction in the relevant jurisdiction.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository
4. Release notes