VoxCPM2Voxcpm
VoxCPM2

VoxCPM2

VoxCPM2 Multilingual Text-to-Speech Voice Design & Controllable Voice Cloning

Text to Speech

Choose a voice, write your script, and generate a clean audio take.

Generate Speech
Enter text in multiple supported languages, and the system will generate speech in the detected language.
0 / 120 characters, 120 remainingEst. 0 credit(s)
🚀
Elon Musk
EN
Deep, conversational and confident male voice. Perfect for online courses and e-learning. Incredible emotional authenticity.
🧽
SpongeBob
EN
Exceptionally energetic and playful voice. Vibrant, child-like male voice perfect for motivational content and sports commentary.
🦅
Donald Trump
EN
Deep, authoritative male voice. Confident, conversational tone perfect for podcasts and voiceovers.
👩
Friendly Women
EN
Warm and friendly female voice. Natural and engaging, perfect for customer service and narration.
🎙️
Energetic Male
EN
Energetic youthful male voice. Great for commercials, social media, and dynamic content.
👩
AD学姐
ZH
自信俏皮活力的女声，语速较快，适合短视频、抖音内容、直播带货。AI TTS技术，声音真实自然。
🐴
赛马娘
ZH
充满朝气、励志且友好的年轻女声。适合短视频、TikTok内容，以及客户服务与虚拟助手。声音富有感染力。
🐺
灰太狼
ZH
青青草原狼大王的雄音，浑厚有力的男声，适合角色扮演和故事叙述。
☯️
太乙真人
ZH
哪吒太乙真人声音，沉稳大气的男性声音，适合神话故事和角色扮演。
🧔
丁真
ZH
年轻温暖的男声，富有叙事感，适合短视频、TikTok内容创作。声音自然流畅。

VoxCPM2 Core Capabilities

Explore the model capabilities behind multilingual TTS, voice design, and controllable voice cloning.

2B Multilingual Speech Model

VoxCPM2 supports text-to-speech in 30 languages and nine Chinese dialects, with a MiniCPM-4 backbone.

Tokenizer-Free Speech Generation

The model avoids external discrete speech tokenizers and generates continuous speech representations; text is still encoded by the language model tokenizer.

Natural-Language Voice Design

Describe a voice in natural language and generate a new voice without providing reference audio.

Controllable Voice Cloning

Provide reference audio and optional style guidance to preserve timbre while adjusting pace, emotion, or expression.

48 kHz Audio Output

VoxCPM2 accepts 16 kHz reference audio and produces 48 kHz output through its AudioVAE pipeline.

VoxCPM2 Model Facts

Official model facts are shown below; online latency and audio quality depend on deployment settings.

2B
Parameters

Latest VoxCPM2 release

30
Languages

Plus 9 Chinese dialects

48 kHz
Audio output

Supports 16 kHz reference input

~0.30
RTF on RTX 4090

Official PyTorch reference

Source: OpenBMB VoxCPM2 model documentation; service results may differ.

VoxCPM Model Versions

Compare official facts for the current VoxCPM2 release and earlier VoxCPM checkpoints.

VoxCPM2
Audio output
48 kHz
Parameters
2B
Reference audio
16 kHz input
Zero-Shot
Multilingual
Open Source
VoxCPM1.5
Audio output
44.1 kHz
Parameters
0.6B
Reference audio
Reference + continuation
Zero-Shot
Multilingual
Open Source
VoxCPM-0.5B
Audio output
16 kHz
Parameters
0.5B
Reference audio
Continuation
Zero-Shot
Multilingual
Open Source

These are official model-version facts; service latency and audio quality depend on deployment.

VoxCPM2 Audio Demos

Explore cross-language speech, emotion, and context-aware generation with the online demo.

Cross-Language - EN→CN

English speaker voice cloned to Chinese speech

Cross-Language - CN→EN

Chinese speaker voice cloned to English speech

Emotion - Happy

Emotionally rich happy tone expression

Emotion - Sad

Emotionally rich sad tone expression

Context-Aware - News

Intelligent news broadcasting style

Context-Aware - Story

Intelligent storytelling style

More audio samples available at Official Demo Page

VoxCPM2 Technical Resources

Read the VoxCPM2 technical report and open-source implementation, then try the official model resources.

Academic Paper

Detailed technical principles, experimental results, and performance evaluation reports

Read Paper

Open Source Code

Complete source code, model weights, and training scripts

Visit Repository

Quick Start

Install the official package and generate multilingual speech with the VoxCPM2 model.

Get Started

Model Download

Pre-trained model weights ready for direct inference

Download Model

API Documentation

Detailed API interface descriptions and usage examples

View Documentation

Community Support

Join the developer community for technical support and discussions

Join Discussion

Technical Highlights

Apache 2.0Open Source License
2BModel Parameters
30Languages
48 kHzAudio Output

Application Scenarios

VoxCPM2 supports practical workflows across content creation, learning, accessibility, and multilingual audio production.

Audiobook Production

Rapidly generate high-quality audiobooks with consistent voice

Language Learning

Personalized speech education with multilingual accent training

Content Creation

Professional voice solutions for video dubbing and podcast production

Accessibility Applications

Personalized reading experiences for visually impaired individuals

Run VoxCPM2 Locally

Install the official package and generate multilingual speech with the VoxCPM2 model.

1

Install the package

Install the official Python package (Python ≥ 3.10, PyTorch ≥ 2.5, CUDA ≥ 12.0).

pip install voxcpm
2

Load the VoxCPM2 model

Load the latest weights from Hugging Face:

from voxcpm import VoxCPM
model = VoxCPM.from_pretrained("openbmb/VoxCPM2", load_denoiser=False)
3

Generate audio

Generate speech and save the native sample rate:

Initialize model
import soundfile as sf
Speech synthesis
text = 'VoxCPM2 generates expressive multilingual speech.'
wav = model.generate(text=text, cfg_value=2.0, inference_timesteps=10, seed=42)
sf.write("output.wav", wav, model.tts_model.sample_rate)

System Requirements

Python 3.10–3.12
PyTorch 2.5+
CUDA 12.0+
See the official deployment guide
About 8GB VRAM for the model reference

Frequently Asked Questions

Common questions and answers about VoxCPM2 technology and usage