When Doraemon Starts Speaking Sichuanese, the Era Where Everyone Can Be a Voice Actor Has Arrived
If you could change your voice, what would you change most?
Language, tone, timbre, emotion — or even conjuring a voice that does not exist in the world at all, purely from imagination. Is that something we can actually achieve?
When VoxCPM 2 can make Doraemon speak Sichuanese with zero human voice actors in the whole process, the answer speaks for itself.
This article includes video content. Please view it in the original WeChat post.
Chinese regional dialects, voice design, general-purpose controllable timbre (especially voice cloning), 48 kHz high audio quality, high expressiveness... Previously, open-source TTS models at home and abroad often needed multiple models to deliver all of these. VoxCPM 2, developed and upgraded by ModelBest together with the OpenBMB open-source community and the Tsinghua University Human-Computer Speech Interaction Lab, integrates them all into a single 2B speech foundation model.
Moreover, while VoxCPM 1 supported only Chinese and English, VoxCPM 2 has greatly improved multilingual performance — covering 30 languages worldwide, including the 8 major languages of Southeast Asia, making expansion into the region far easier.
Take ModelBest's "Little Cannon" ad film, for example: VoxCPM 2 effortlessly produced voiceovers in multiple Southeast Asian languages, saying goodbye to robotic delivery while sounding genuinely human and down to earth:
One 2B model handles video creation, podcasts, dubbing, audiobooks, going global, and the voice needs of every kind of internet creator — and it is open-source and free.
Four Abilities in One Model
For speech foundation models, ModelBest likewise follows the "Densing Law" of efficient training, aiming for the highest possible knowledge density at the smallest possible parameter scale. At just 2B parameters, VoxCPM 2 achieves remarkable results in speech generation.
It brings multilingual support, voice design, general-purpose timbre control, and high audio quality and expressiveness into one model, opening up broad room for developers and AI voice enthusiasts.
01 Global Reach: 30 Languages + 9 Major Dialects
VoxCPM 2 supports 30 major languages worldwide, with especially thorough R&D upgrades for Southeast Asian languages, covering the 8 major languages of the region: Vietnamese, Thai, Indonesian, Lao, Burmese, Khmer, Filipino, and Malay.
How does it sound? Here are some examples:
The famous blood-test scene from the Southeast Asian version of Empresses in the Palace, switching smoothly between Thai and Vietnamese:
Beyond its international flair, VoxCPM 2's training also reflects the distinctive perspective of a Chinese large-model team: it has mastered 9 major Chinese dialects — Sichuanese, Cantonese, Wu, Northeastern, Henanese, Shaanxi, Shandong, Tianjin, and Hokkien — so your voice AI can be not just "Tony" but also "Dazhuang," not just "Lisa" but also "Cuihua."
VoxCPM 2 with a hometown flavor is already being put to wildly creative use:
Henanese: It's freezing today! Let's hurry back and have a bowl of soup, huh?
Shanghainese: I don't understand why everyone is talking about Xiang Yu being trapped at Gaixia...
Hokkien: A young, sweet-voiced Hokkien woman recalls life back home
Cantonese: I never felt like I was a left-behind child...
Beyond its international flair, VoxCPM 2's training also reflects the distinctive perspective of a Chinese large-model team: it has mastered 9 major Chinese dialects — Sichuanese, Cantonese, Wu, Northeastern, Henanese, Shaanxi, Shandong, Tianjin, and Hokkien — so your voice AI can be not just "Tony" but also "Dazhuang," not just "Lisa" but also "Cuihua."
02 Versatile Voice Actor: Voice Design, Where Text Becomes Sound
With VoxCPM 2, if you don't want to reveal a real person's voice and can't find the right voice actor, you can create your own one-of-a-kind "versatile voice actor" from imagination alone — no reference audio needed.
All it takes is a few keystrokes: enter a text description in VoxCPM 2 specifying the timbre, emotion, gender, age, and more of the voice you imagine, and VoxCPM 2 conjures a brand-new voice out of thin air.
Have VoxCPM 2 generate 7 voices in one go and let them star in a wuxia blockbuster:
Reciting Li Bai's "Bring in the Wine" with bold, soaring spirit:
Performing Du Fu's "On the Height" in a grave, halting tone:
03 Every Voice Different: General-Purpose Timbre Control
Most speech models at home and abroad support control over only a few fixed timbres when cloning voices. To break this limit, VoxCPM 2 takes the unconventional route of diffusion autoregressive continuous representation, achieving general-purpose timbre control in the true sense.
Compared with conventional token-based methods, the diffusion autoregressive architecture preserves more acoustic information, so voice cloning is no longer a stiff pastiche but a highly faithful regeneration of speech. Emotional swells, shifts in mood, even subtle breaths are reproduced with maximum similarity.
The Queen of the Women's Kingdom in Journey to the West:
Original voice:
VoxCPM 2:
The char siu segment from A Bite of China:
Original voice:
VoxCPM 2:
04 Cinema-Grade Dubbing: High Audio Quality, High Expressiveness
Sample rate is the most direct measure of audio quality: 8000 Hz is enough only for phone calls, 16000 Hz delivers merely clear audio, while VoxCPM 2 goes straight to 48000 Hz, reaching Hi-Fi level!
With its high audio quality, AI-generated speech can be used in a wider range of fields with higher audio demands, such as film and TV dubbing. The emotional impact of the voice is also greatly enhanced.
The English version of Every Treasure Tells a Story:
With high-quality generation and reproduction, the emotion in the voice stands up to close scrutiny, putting listeners right there:
A warm, very fast-speaking male voice; in Chinese it means: Bro, spicy or sweet? Here you go, extra-crispy puchka! Want some more? Ma'am, I made this one extra spicy just for you. Here, twenty rupees!
A rough, hoarse male voice with heavy rolled consonants, full of menace and pressure: Don't you talk big to me here! Are you paying back the money or not? Say it clearly, right here, right now! Don't think there'll be a next time, you bastard!
Not Just Capable — Free and Open-Source!
VoxCPM 2 is not here to replace anyone; it is here to empower everyone who needs it.
To that end, ModelBest remains committed to open source and has worked hard to be developer-friendly.
VoxCPM 2 supports native Torch inference, full-parameter fine-tuning, and LoRA fine-tuning. Whether you are an enterprise with top-tier compute or an individual beginner with a beat-up laptop, a mature ecosystem supports one-click runs.
VoxCPM 2 is now fully available and open-sourced. We invite developers and creators worldwide to open a new era of high-fidelity speech together.
Ready to make something big with voice? Click the open-source, demo, and download links below to build a voice world of your own!
➤ Demo
🔗 https://voxcpm.modelbest.cn/
➤ Hugging Face
🔗 https://huggingface.openbmb.com/model/openbmb/VoxCPM2
➤ GitHub
🔗 https://github.com/OpenBMB/VoxCPM/
➤ ModelScope
🔗 https://modelscope.cn/models/OpenBMB/VoxCPM2
➤ GitCode
🔗 https://ai.gitcode.com/OpenBMB/VoxCPM2
➤ Modelers
🔗 https://modelers.cn/models/OpenBMB/VoxCPM2