Reverb offline, voice clone, 7 languages, non-commercial

VoicesOfflineCPU onlyVoice cloningNon-commercial

Dubs your subtitles in English, German, French, Spanish, Italian, Polish and Portuguese with Reverb-TTS, by Software Mansion AI Labs, in one of 16 voices or in each speaker's own voice cloned from a few seconds of the video and its subtitles. It runs on your own computer's processor, with no graphics card. The model is for non-commercial use only.

Install it from Subtitld

Settings Add-ons Reverb (offline, voice clone, 7 languages, non-commercial) Install

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

Source on GitHub
Languages
7
Models
3
Voices
17
Add-on download
40–65 MB
RAM, at least
2.5 GB

What it does

Seven languages

English, German, French, Spanish, Italian, Polish and Portuguese. Numbers, dates, prices and units are read out as words, in each language's own way.

The speaker's own voice

Subtitld cuts a few seconds of each speaker from the video and sends them with the subtitles spoken in them. Reverb clones that voice in any of its seven languages. Or pick one of 16 voices.

Exact or Fast

Exact runs the model as it was published. Fast is about twice as fast and a smaller download, at a small cost in accuracy.

Offline after the first line

The model downloads once, the first time a line needs it. After that, nothing leaves your computer.

Non-commercial model. The Reverb-TTS weights and its preset voices are under CC-BY-NC-4.0, which doesn't permit commercial use. The add-on's code is Apache-2.0, but that doesn't change the weights' terms. For commercial work, ZONOS3-Mini also clones voices in all seven of these languages, and its weights are Apache-2.0.

Before you install

  • Disk: 640 MB for the model, downloaded on the first line; 378 MB instead if you choose Fast before the first line. The first time you use a speaker's own voice, 327 MB more.
  • Memory: about 1.5 GB while it speaks.
  • Time: on an Intel Core i7-13700T that was busy with other work, a second of speech took about 0.8 seconds to make in Exact speed and about 0.6 seconds in Fast.
  • Processor: a 64-bit Intel or AMD processor, or Apple Silicon. No graphics card is used.

Voices

In the Speakers tab, choose Reverb 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.

17 voices.

Recorded in Women Men
English Narrator, whispering Narrator 1, narrator 2, angry
French French woman French man
German German woman German man
Italian Italian man, expressive Italian man
Polish Polish man
Portuguese Portuguese woman Portuguese man
Spanish Spanish man 1, Spanish man 2

These are Reverb's own preset voices. Every voice speaks every language, but with the accent of the language it was recorded in. For a language with no voice of its own here, or no woman's voice, use the speaker's own voice.

With Speaker's own voice, Subtitld cuts a sample of that speaker from the project's audio, from one of their subtitles or several short ones joined, and sends what is said in it: the text of those subtitles. Reverb needs both. It continues from the sample and its words, and without the words it would drop the first words of every line.

So the subtitles must be in the language that is spoken in the video. If you translated them, Subtitld sends the original text, as long as it knows which language the audio is in: set it under Subtitle Language in the Transcription panel. Otherwise the line stops with a message that says so. Subtitles that paraphrase what is said, rather than transcribe it, make the clone worse.

The best sample is 3 to 10 seconds of one person speaking, clean and close to the microphone. A sample shorter than 2 seconds is refused, and one longer than 15 seconds tends to give worse lines. The Fallback reference clip in the add-on's settings, with its Fallback clip's transcript, is used only when no sample can be cut.

Each line is made with the seed Subtitld sends, so the same line in the same voice comes out the same every time.

Models

The model is not inside the add-on: it downloads into Subtitld's shared model cache the first time a line needs it, from the add-on's own release on GitHub, where Reverb's weights are kept converted for ONNX Runtime, each file checked. An interrupted download resumes where it stopped. Updating or removing the add-on keeps it.

Model DownloadGood for
Reverb-TTS v1.0.0 (ONNX, full precision) with the Higgs Audio v2 decoder and 16 preset voices Default reverb
640 MB
Every line in Exact speed, the default: the network, the decoder and the 16 voices
Reverb-TTS v1.0.0 network in int8 (Fast mode) reverb-fast
266 MB
Every line in Fast speed, in place of the Exact network: 378 MB in all with the decoder
Higgs Audio v2 encoder (voice cloning) reverb-codec-encoder
327 MB
The speaker's own voice, the first time you dub one

Exact and Fast share the decoder and the voices (111 MB): only the network differs, 529 MB for Exact and 266 MB for Fast. Switching from one to the other later downloads the other network the first time a line needs it.

The add-on runs Reverb on ONNX Runtime, without PyTorch. Reverb's sampler and text preparation are ported from Software Mansion's own code. Its random numbers come from a different source, so a seed gives a different take here than in Reverb's own package; the Exact network is checked to predict what that package predicts.

Options

In Settings › Add-ons › Reverb (offline, voice clone, 7 languages, non-commercial) › Configure.

CPU threads
Default: 0

0 uses the computer's logical processors, at most 8. More is rarely faster; fewer keep the computer more responsive while it dubs.

Quality steps
Default: 8

How many passes the model makes over each line. 8 is Reverb's own setting; fewer is faster, with more slips.

Speed
Default: exact

Exact runs the model as Software Mansion published it. Fast runs it in 8-bit integers, about twice as fast, with its own 266 MB download: its lines usually sound the same, but it strays further from the published model and slips more often.

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 10 seconds of one person speaking, clean.

Fallback clip's transcript
Default: empty

Exactly what is said in the fallback clip, word for word. Reverb clones from the clip and its words, so the fallback clip is refused without it.

Languages

English, German, French, Spanish, Italian, Polish and Portuguese. A project in any other language gets a message that says so: pick another dubbing engine for it.

  • en
  • de
  • fr
  • es
  • it
  • pl
  • pt

Platforms

System DownloadTested
Linux x86-64 64.9 MBSpoke all seven languages and cloned a voice with the real model in CI
macOS (Apple silicon) 45.6 MBSpoke all seven languages and cloned a voice with the real model in CI (Apple Silicon)
Windows x86-64 40.5 MBSpoke all seven languages and cloned a voice with the real model in CI

Linux x86-64 (glibc 2.35 or newer, as in Ubuntu 22.04, Debian 12 or Fedora 36), macOS on Apple Silicon and Windows x86-64. Each release is built on all three, and on each one CI installs the add-on as Subtitld does, downloads the model, speaks a line in each language and clones a voice.

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: 640 MB (378 MB in Fast), and 327 MB more for the speaker's own voice.
  • unsupported_languageThe project's language is not one of the seven, or not set. Set it in the project, or pick another dubbing engine.
  • unsupported_voiceThe project names a voice this version doesn't have. Pick the voice again in the Speakers tab.
  • bad_paramsThe line has no text, or the cloned voice got no sample or no transcript of it. For a translated project, set Subtitle Language to the language spoken; with no video open, set a Fallback reference clip and its transcript. The message says which.
  • out_of_memoryThe computer ran out of memory. Close other programs: Reverb needs about 1.5 GB free.
  • internalSomething failed inside the add-on; the message has the details.

License

Add-on
Apache-2.0
with Reverb's own code, ported to numpy
Engine
MIT
ONNX Runtime
Model
CC-BY-NC-4.0
Reverb-TTS v1.0.0 weights, by Software Mansion AI Labs; non-commercial only
Voices
CC-BY-NC-4.0
Reverb's 16 preset voices; non-commercial only
Text reading
MIT, Unicode-3.0
Verbalis, by Software Mansion, with a table from the Unicode Character Database

The add-on is Apache-2.0, and so is Reverb's code that it ports. The Reverb-TTS weights and its 16 preset voices are CC-BY-NC-4.0: non-commercial use only. A license file that comes with the model says what was changed in converting it. ONNX Runtime is MIT, and Verbalis, which reads numbers and dates, is MIT.

Use it responsibly. A cloned voice sounds like a real person. Clone only voices you have the right to use, with the speaker's consent; don't impersonate real people, and don't use the speech you make to deceive, defraud or harass. Tell your listeners when they hear a synthetic voice.

Versions

  • 26.11.1

    Text-to-speech for Subtitld with Reverb-TTS v1.0.0 (Software Mansion AI Labs): English, German, French, Spanish, Italian, Polish and Portuguese at 24 kHz, in 16 preset voices or in the speaker's own voice, cloned from a clip of the video and its subtitles.

    The archive holds the add-on, ONNX Runtime, tokenizers and Verbalis. The model (640 MB, or 378 MB in Fast mode) downloads on the first line, and the codec encoder (327 MB) on the first cloned line, from this repository's models-1 release and Reverb's own repository. Every file is pinned and checked by sha256. It runs on the CPU.

    The model is licensed under CC BY-NC 4.0: non-commercial use only. The add-on is Apache-2.0, and so is Reverb's code it ports. 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 reverb-<version>-libquadmath-source.tar.gz in this release.

Other Voices add-ons