VoxCPM2

VoxCPM2 is a text-to-speech model for turning written text into spoken audio. The practical angle is that it is not limited to one fixed narrator. You can give it normal text, describe the kind of voice you want, or provide a short reference clip when you need voice cloning.

That makes the workflow useful for more than a demo sentence. For example, you can write a short product walkthrough, ask for a calm older narrator or a faster upbeat delivery, then compare whether the generated voice actually fits the script. If you have a permitted reference voice, you can also test whether style instructions change pace, emotion, or emphasis without losing the speaker's basic timbre.

The model page also describes 30-language support, 48kHz output, streaming generation, and local examples through the voxcpm package. The public Space is the fastest way to judge whether the extra voice control is useful before you spend time on a CUDA setup.

Start with the Hugging Face Space and run one short script twice. First use plain text-to-speech. Then add a voice description such as age, tone, pace, or emotion and listen for whether the delivery changes in a useful way instead of only sounding different.

For a second test, try a language you actually need. The model supports 30 languages, but quality can vary by language and training coverage, so do not judge it only from English or Chinese examples.

If the browser demo is promising, move to the model page for local use. The quick start installs voxcpm, loads openbmb/VoxCPM2, and writes generated audio with Python. Treat local use as a GPU workflow: the model card lists Python 3.10 or newer, PyTorch 2.5 or newer, CUDA 12 or newer, and about 8GB of VRAM.

Voice cloning needs a stricter bar than ordinary text-to-speech. Only use voices you have permission to use, label synthetic audio where appropriate, and test the output for unstable emphasis, language-specific quality gaps, or odd behavior on very long and expressive text.

• Prototype narrated product demos, explainers, tutorials, or internal training audio.
• Compare voice descriptions before choosing a narrator style for a project.
• Test multilingual voiceover drafts without starting from a closed speech API.
• Clone a consented reference voice and steer pace, emotion, or delivery for controlled experiments.
• Build voice-agent or interactive audio prototypes that need streaming speech.

Try the demo

View model page