VoxCPM — Multilingual Voice Cloning Without a Tokenizer
OpenBMB's VoxCPM2, running in the project's own Hugging Face Space
Four things are worth testing here: multilingual synthesis, a voice built from a written description, cloning with expression controls on top, and how 48 kHz output actually sounds. Everything on this page is our own summary of public VoxCPM2 material, wrapped around the official demo.
The panel above is the VoxCPM-Demo Space on Hugging Face, published by OpenBMB. The first generation is slow while the hosted demo wakes; later ones are not. Keep confidential text and private voice samples out of any public Space.
What VoxCPM is
An open TTS family from OpenBMB. What sets the current VoxCPM2 release apart is structural: it does without the discrete speech-tokenizer stage most systems rely on, and aims that architecture at natural multilingual speech, designed voices, and convincing cloning.
The model card and project docs put it at 2B parameters, tokenizer-free and diffusion autoregressive, trained on more than 2 million hours of multilingual speech. Thirty languages are supported, a short reference clip is enough for cloning, and audio comes out at 48 kHz through AudioVAE V2.
The Space is worth the five minutes because it surfaces exactly the four things a team checks first: plain multilingual synthesis, a voice described in words, cloning with style guidance layered on, and the higher-fidelity route that takes a reference clip together with its transcript.
What makes it different
Multilingual generation, designed voices, cloning, and a serving path meant for production — in one open release.
Tokenizer-free diffusion autoregressive TTS
Continuous speech representations come straight out of the diffusion autoregressive stack — no discrete token layer in between.
30-language multilingual speech
Thirty languages on the model card — Chinese, English, Japanese, Korean, Spanish, French, German, Portuguese, Hindi, Arabic and Vietnamese among them.
Describe the voice you want
Gender, age, tone, emotion, pace — write the speaker down and VoxCPM2 constructs one, no reference recording involved.
Two levels of cloning
A short clip preserves timbre; style guidance on top steers the delivery; adding the clip's transcript unlocks the higher-fidelity continuation route.
48 kHz out, and a streaming route
References come in at 16 kHz, speech goes out at 48 kHz. The docs describe real-time streaming under standard inference, with Nano-vLLM serving faster still.
Open, with conditions attached
Apache-2.0 on the licence, but the project draws an explicit line at impersonation, fraud and disinformation. Label what you generate.
Four steps to a sample
The hosted Space covers the main workflows with nothing installed locally.
Start the Space
Work in the embedded panel above, or pop it into its own tab. A cold start takes a while — dependencies and weights both have to load.
Paste text in your language
No language tag needed — VoxCPM2 takes multilingual input as it comes. Begin with one short sentence in the language you actually care about.
Design a voice, or clone one
Designing: write the speaker description ahead of your text. Cloning: upload a clean clip, and only reach for style guidance if you want the pace or emotion changed.
Iterate on a short sample
Inference steps, guidance, denoise and style are all adjustable. Generate something short, listen for pronunciation, pacing and speaker likeness, then change one setting at a time.
The specification in one table
Drawn from the model card, Space metadata, documentation and demo page, condensed so you can scan it.
Where it fits
The combination that matters: many languages, expression you can steer, and the freedom to run it on your own hardware.
Narration in several languages
Draft the voiceover for a product tour, a lesson or an explainer in each target language before anyone books a human read.
Prototyping an agent voice
Try personas, pacing and emotional register — and hear what 48 kHz buys you — before committing to a production voice stack.
Dubbing and localisation checks
Test pronunciation, rhythm and expressiveness on localised scripts, particularly in the languages older TTS systems handled badly.
Cloning, with consent first
Where you have the speaker's permission, cloning plus style guidance lets you vary emotion and pacing while the timbre stays put.
Inventing a voice
Voice design is the route to a fictional or brand-safe speaker without any real person being involved.
Hands-on evaluation
Benchmark it against other open TTS, run the Python package, or dig into the Nano-vLLM serving path if latency is what you are chasing.
Getting better output
- Judge pronunciation, rhythm and consistency on a sentence before you commit to a paragraph.
- Clone from clean single-speaker audio — noise or a second voice makes similarity impossible to assess fairly.
- A designed voice that misses usually needs a more concrete description: age, pace, emotion, vocal texture.
- Supply the reference transcript when asked — that alignment is what preserves the original delivery in high-fidelity cloning.
- Neither confidential scripts nor private voice samples belong in a public Space; self-host VoxCPM2 for that work.
- Do not use VoxCPM for impersonation, fraud or unlabelled synthetic media. Consent comes before any real voice is cloned.
VoxCPM questions, answered
For teams weighing it up for multilingual synthesis, cloning, designed voices, or a self-hosted voice stack.
What VoxCPM is
OpenBMB's tokenizer-free TTS family. The current release, VoxCPM2, is a 2B-parameter model covering multilingual synthesis, designed voices and cloning.
Can I try VoxCPM online?
Yes — the official VoxCPM-Demo Space is embedded above, so VoxCPM2 runs in your browser with no Python package installed.
What languages does VoxCPM2 support?
Thirty, per the model card — Chinese, English, Japanese, Korean, Spanish, French, German, Portuguese, Hindi, Arabic and Vietnamese among them — with several Chinese dialects listed separately.
Does VoxCPM support voice cloning?
Two ways. Controllable cloning works from reference audio alone; the higher-fidelity path uses the clip together with its transcript. Either way, restrict it to voices you own or have permission to use.
What is VoxCPM voice design?
You describe the speaker in plain language — age, gender, tone, emotion, pace — and get speech in a voice built to that description, with no recording uploaded at all.
Is the embedded VoxCPM demo private?
It is not. That Space is public and runs outside our infrastructure. Confidential text and private voice samples should not go into it — self-host for anything sensitive.
Hear it before installing it
Synthesise in a second language, describe a voice into existence, and try cloning with a voice you are permitted to use — all before VoxCPM touches your own machine.