Download

VoxCPM2 (offline, voice clone, 30 languages)

VoicesOfflineGPU optionalVoice cloningCommercial use OK

Dubs your subtitles in 30 languages with OpenBMB's VoxCPM2, in each speaker's own voice cloned from the video, or in a voice made from a description. It runs on your own computer, and the add-on, the code and the weights are all Apache-2.0.

Install it from Subtitld

Settings Add-ons VoxCPM2 (offline, voice clone, 30 languages) Install

Subtitld downloads it, checks its SHA-256, and keeps it updated.

Source on GitHub
Languages
30
Models
1
Voices
9
Add-on download
205–318 MB
RAM, at least
7 GB

What it does

The speaker's own voice

Subtitld cuts a clip of each speaker from the video and sends it with the words spoken in it. VoxCPM2 continues from that clip, so the dub sounds like the person on screen, in any of its 30 languages.

Or a voice you describe

Seven ready-made voices, from a warm woman to a gravelly older man and a narrator, or your own words: "a young man, fast and cheerful".

Studio sound

Speech comes out at 48 kHz, the rate video uses.

Free for commercial work

The add-on, VoxCPM's code and the VoxCPM2 weights are all Apache-2.0.

Before you install

VoxCPM2 is a large model, 2.4 billion parameters. It gives the most natural dubs of the offline add-ons, and asks the most of your computer:

  • Disk: about 5.6 GB, most of it the 4.96 GB model, downloaded on the first line you dub.
  • Memory: about 7 GB free. With less it still runs, but far slower.
  • Time: on a CPU each second of speech takes several seconds to make: 5 to 6 seconds on an Intel Core i7-13700T. A ten-minute video can take about an hour. On an Apple Silicon Mac it runs on the GPU.

For quick dubs on modest computers, Piper or Kokoro are much lighter.

Voices

In the Speakers tab, choose VoxCPM2 as a speaker's Dubbing engine, then pick a Voice. Turn on Fit dub to subtitle duration for the speaker to stretch or squeeze each line to its subtitle.

9 voices.

The cloned voice

With Speaker's own voice, Subtitld cuts a sample of that speaker from the project's audio, from one of their subtitles of 3 to 12 seconds, or several short ones joined. If you have separated music from voice for this video, the sample comes from the voice alone. The Fallback reference clip in the add-on's settings is used only when no sample can be cut.

Subtitld sends the subtitle text with the sample, and VoxCPM2 carries on from it, which comes closest to the speaker, delivery included. If a dub into another language keeps the speaker's accent, turn on Clone the timbre only.

Described voices

VoxCPM2 makes up a new voice every time it reads a description. So the add-on makes each speaker's voice once, and every line of that speaker reuses it: the voice stays the same through the video, and two speakers given the same description still sound like two people.

Models

There is one model, VoxCPM2. It is not inside the add-on: it downloads into Subtitld's shared model cache the first time you dub, from Hugging Face at a pinned revision, each file checked. An interrupted download resumes where it stopped. Updating or removing the add-on keeps the model.

Model DownloadGood for
VoxCPM2 (2.4B) Default voxcpm2
4.7 GB
Every voice, in all 30 languages

The first line also loads the model, so it takes longer than the ones after it. Cancel stops a line between two steps, or stops the download.

Options

In Settings › Add-ons › VoxCPM2 (offline, voice clone, 30 languages) › Configure.

Inference device
Default: auto

Automatic uses the GPU of an Apple Silicon Mac, otherwise the CPU. This build does not use NVIDIA graphics cards.

Weight precision
Default: auto

Automatic suits almost every computer. float32 is a little faster but needs about 11 GB of memory.

Quality steps
Default: 10

More steps, clearer speech, slower. 5 is about a quarter faster than 10, with audibly less clarity.

Guidance
Default: 2

How closely the speech follows the text and the voice. VoxCPM's default is 2.0.

CPU threads
Default: 0

0 picks the performance cores on CPUs that also have efficiency cores. Fewer threads keep the computer more responsive while it dubs.

Voice description
Default: empty

Used by the voice "Voice you describe": who speaks and how, in a few words, for example "An elderly woman, soft and warm voice, slow pace". English works for every language.

Fallback reference clip
Default: empty

A clip for the cloned voice when Subtitld can't cut one from the video, for example with no video open: 3 to 15 seconds of one person speaking.

Fallback reference transcript
Default: empty

The exact words spoken in the clip above. With them the clone comes closer to the speaker.

Clone the timbre only
Default: off

Clones only the sound of the voice, not the clip's delivery. Try it when a dub into another language keeps the speaker's accent, or when your subtitles paraphrase the speech.

Languages

Arabic, Burmese, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Khmer, Korean, Lao, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, Tagalog, Thai, Turkish and Vietnamese. VoxCPM2 reads the language from the text itself, so every voice speaks every language.

  • ar
  • my
  • zh
  • da
  • nl
  • en
  • fi
  • fr
  • de
  • el
  • he
  • hi
  • id
  • it
  • ja
  • km
  • ko
  • lo
  • ms
  • no
  • pl
  • pt
  • ru
  • es
  • sw
  • sv
  • tl
  • th
  • tr
  • vi

Platforms

System DownloadTested
Linux x86-64 312.0 MBDubbed with the real model on a real machine; built and checked in CI
macOS (Apple silicon) 318.0 MBBuilt and checked in CI (Apple Silicon)
Windows x86-64 204.6 MBBuilt and checked in CI

Linux x86-64, macOS on Apple Silicon and Windows x86-64. Each release is built on all three, and CI checks that everything needed to speak is in the build. CI doesn't run the model itself.

Troubleshooting

  • network_unavailableThe model couldn't be downloaded. Check your connection and try again: the download resumes.
  • disk_fullNot enough disk space for the model. It needs about 5 GB free.
  • oomOut of memory. Close other programs; VoxCPM2 needs about 7 GB free, 11 GB in float32.
  • bad_paramsThe line has no text, or the cloned voice got no sample: open the video, or set a Fallback reference clip. The message says which.
  • unsupported_voiceThe project names a voice this version doesn't have. Pick the voice again in the Speakers tab.
  • internalSomething failed inside the add-on; the message has the details. If you changed Inference device, set it back to Automatic.

License

Add-on
Apache-2.0
Engine
Apache-2.0
VoxCPM, by OpenBMB
Model
Apache-2.0
VoxCPM2 weights

Fine for commercial work: the add-on, VoxCPM and the VoxCPM2 weights are all Apache-2.0. The dubs you make are yours. A cloned voice sounds like a real person: clone voices you have the right to use, and say that a dub is synthetic where it matters.

Versions

  • 0.0.1

    Text-to-speech for Subtitld with OpenBMB's VoxCPM2: 30 languages at 48 kHz, in the speaker's own voice cloned from the video, or in a voice made from a description.

    The archive holds the add-on and a CPU build of PyTorch. The 4.96 GB model downloads on the first synthesis, at a pinned revision, checked by sha256. Plan for about 7 GB of free memory; on a CPU each second of speech takes several seconds to make.

    The add-on, VoxCPM and the VoxCPM2 weights are Apache-2.0. Each archive carries THIRD_PARTY_LICENSES/ with the licence of everything frozen into it. The Linux archive carries libquadmath (LGPL-2.1, with numpy); its complete source is voxcpm-<version>-libquadmath-source.tar.gz in this release.

Other Voices add-ons

Updated · From the add-ons catalog Report a problem with this add-on