VoxCPM2

VoxCPM2

Speaks text in 30 languages and nine Chinese dialects at 48 kHz, designs a new voice from a written description and clones a voice from a short clip. An open 2.3B text-to-speech model for developers.

Made by openbmb on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: VoxCPM2 is an open text-to-speech model from openbmb that works without a tokenizer and uses a diffusion autoregressive design. The Hugging Face record lists 2.3B parameters, which the card rounds to 2B. It speaks 30 languages plus nine Chinese dialects, writes 48 kHz audio and was trained on more than 2 million hours of speech. Besides plain reading, it can invent a voice from a written description and can copy a voice from a short sample.

Who it suits: Developers and creators who need multilingual speech, custom voices or voice cloning, and who have an NVIDIA GPU to run it.

Industry | Who uses it | How they use it | Example
Media and video | Creators | Narrate scripts in many languages | A Spanish voice-over for a short film
Gaming and animation | Studios | Design a character voice from a description | A gentle young voice for a guide character
Customer service | Voice agent builders | Stream spoken replies with low delay | A phone bot that answers in French
Accessibility | Tool makers | Read text aloud in the listener's language | A Hindi reader for a news app

If you have no NVIDIA GPU, you can skip it, because the card lists CUDA 12 or later as a requirement.

Where to use it:

• Python, through the voxcpm package
• A live demo and docs that the card points to
• A faster serving option, Nano-VLLM, that the card mentions for streaming

Use case | Who | What they get back
Plain speech | Developers | A 48 kHz audio file of your text
Voice design | Creators | A new voice from a written description, with no sample needed
Voice cloning | Producers | Speech in the voice of a clip you supply, with optional style hints

How to set it up:

• Python 3.10 or later
• PyTorch 2.5.0 or later
• CUDA 12.0 or later on an NVIDIA GPU
• About 4.6 GB of disk space in BF16, or about 1.1 GB at 4-bit (an estimate)

1. Install the voxcpm package.
2. Download the model files, or let the package fetch them on first load.
3. Load the model in Python and ask it to speak.

pip install voxcpm
hf download openbmb/VoxCPM2

How to use it: Load the model in Python and call its generate method with your text, then save the audio to a file. For voice design, put a short description of the voice in brackets before the words to speak. For cloning, pass a reference clip of the speaker. For the closest copy, also pass the clip's exact transcript. A streaming method returns audio in chunks as it is made.

Example: You write a sentence with the description of a young woman with a gentle, sweet voice at the start. The model speaks the sentence in a voice that fits that description and saves it as a 48 kHz file, with no reference audio.

For newcomers and experts: If you are new to this, start with plain speech, then try voice design, which needs no sample. Only clone a voice that is yours or that you have the right to use.

It accepts a 16 kHz reference clip and raises the output to 48 kHz inside its own audio codec, so no separate upsampler is needed. The card gives a real-time factor of about 0.3 on an RTX 4090, and about 0.13 with Nano-VLLM.

You write text, with an optional voice note
You add a reference clip to clone a voice
The model generates the speech
You save or stream 48 kHz audio

Good to know:

• Languages: 30 languages, with no language tag needed in the text
• Voice cloning: use only voices you have the right to use
• Variants: the Hugging Face record lists 22 quantized and 50 fine-tuned versions made by others

Title: Real-time factor, where lower is faster
Plain run on an RTX 4090: 0.3 RTF
With Nano-VLLM: 0.13 RTF

• License: apache-2.0, free for commercial use according to the card
• Cost to run: no API key or paid service is named, but you need your own NVIDIA GPU
• Data it reaches: downloading fetches model files, and the model reads your text and any reference clip you give it
• Not for: cloning someone's voice without the right to use it

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: voxcpm
• Parameters: 2.3B
• Weights: about 4.6 GB in BF16; about 1.1 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 984,860
• Likes: 1,783
• License: apache-2.0
• Languages: zh, en, ar, my, da, nl, fi, fr
• Quantized versions: 22
• Fine-tunes: 52
• Tags: voxcpm, safetensors, text-to-speech, multilingual, voice-cloning, voice-design, diffusion, audio

Links

• Model: https://huggingface.co/openbmb/VoxCPM2
• Files: https://huggingface.co/openbmb/VoxCPM2/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download openbmb/VoxCPM2

Then follow the usage in its model card. Nothing in this kit runs by itself.