ZONOS3-Mini offline, voice clone, 43 languages

VoicesOfflineCPU onlyVoice cloningCommercial use OK

Dubs your subtitles in 43 languages with ZONOS3-Mini, by Zyphra, in one of 27 voices or in each speaker's own voice cloned from a few seconds of the video. It runs on your own computer's processor, about twice as fast as real time, with no graphics card.

Install it from Subtitld

Settings Add-ons ZONOS3-Mini (offline, voice clone, 43 languages) Install

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

Source on GitHub
Languages
43
Models
2
Voices
28
Add-on download
19–24 MB
RAM, at least
2 GB

What it does

43 languages

From Arabic, Bengali and Bulgarian to Turkish, Urdu and Vietnamese, including Japanese, Korean, Mandarin, Hindi and eight more languages of India. Numbers, dates, prices and units are read out in each language's own grammar.

The speaker's own voice

Subtitld cuts a sample of each speaker from the video and sends it with every line. ZONOS3-Mini clones that voice in any of its languages. Or pick one of 27 voices.

Lines end where the text ends

Each line is checked while it is spoken, so words aren't skipped, and it stops when the text is read instead of carrying on with words that aren't in the subtitle.

Offline after the first line

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

Before you install

  • Disk: 1,115 MB for the model, downloaded on the first line, and 45 MB more the first time you use a speaker's own voice.
  • Memory: about 1.6 GB while it speaks.
  • Time: on an Intel Core i7-13700T, a second of speech takes about half a second to make. The first line waits about 4 seconds more while the model loads, and the first line in a speaker's own voice about 5 seconds more while the voice is learned.
  • Processor: a 64-bit Intel or AMD processor with AVX2 (Intel since 2013, AMD since 2015), or Apple Silicon. No graphics card is used.

Voices

In the Speakers tab, choose ZONOS3-Mini 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.

28 voices.

Recorded in Women Men
English Imogen (British), Morgana (witch), Sakura (anime), Savannah (American), Willow (ASMR) Cillian (Irish), Jett (energetic), Rupert (British), Thaddeus (announcer), Wyatt (American)
French Gaspard
German Liesel Wolfgang
Hindi Ananya Arjun
Italian Ginevra Lorenzo
Korean Seoyeon Minjun
Mandarin Meilin (energetic) Haoran (energetic)
Portuguese Leonor Thiago
Spanish Ximena Rodrigo
Telugu Sravya Chaitanya

These are Zyphra's own default voices. Every voice speaks every language, but with the accent of the language it was recorded in, and in another language that accent can be strong: an English voice speaking Russian is hard to follow. For a language with no voice of its own here, 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. 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.

ZONOS3-Mini listens to up to a minute of the sample and copies the voice, not the words or their delivery. The best sample is clean speech close to the microphone: noise and echo make the clone less like the speaker, and a phone recording gives a phone-sounding voice.

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 Zyphra's Hugging Face repository at a pinned revision, each file checked. An interrupted download resumes where it stopped. Updating or removing the add-on keeps it.

Model DownloadGood for
ZONOS3-Mini (backbone Q4/Q8 + waveform decoder, GGUF) Default zonos3-mini
1.1 GB
Every line, in every language and voice
ZONOS3-Mini speaker encoder (voice cloning) zonos3-speaker-encoder
45 MB
The speaker's own voice, the first time you dub one

The add-on runs Zyphra's own ggml runtime, the same code as Zyphra's local server, built into the add-on without PyTorch.

Left alone, the model sometimes keeps talking after a short line, adding a few words that aren't in the subtitle. So every line runs under LACI, Zyphra's alignment check: it follows which part of the text is being spoken, takes a skipped or stuck passage again, and stops the line once the text has been read. The add-on then cuts the line at the first silence after that point. That cut is the add-on's own rule, not Zyphra's. In our test of short subtitle lines, 61 of 64 came out clean this way, against 37 of 64 without. A line that comes out silent is made again.

Options

In Settings › Add-ons › ZONOS3-Mini (offline, voice clone, 43 languages) › 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.

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: 5 to 20 seconds of one person speaking, clean, not a phone recording.

Languages

Arabic, Bengali, Bulgarian, Chinese (Mandarin), Czech, Danish, Dutch, English, Finnish, French, Georgian, German, Greek, Gujarati, Hebrew, Hindi, Hungarian, Indonesian, Italian, Japanese, Kannada, Korean, Malay, Malayalam, Marathi, Norwegian, Odia, Polish, Portuguese, Punjabi, Romanian, Russian, Slovak, Spanish, Swedish, Tagalog, Tamil, Telugu, Thai, Turkish, Ukrainian, Urdu and Vietnamese. A project in any other language gets a message that says so: pick another dubbing engine for it.

English follows the project's region: the United Kingdom, Ireland, Australia, New Zealand, South Africa and India get British English, any other region American English. Portuguese is read as Brazilian Portuguese, and French as the French of France.

  • en
  • fr
  • de
  • es
  • it
  • pt
  • ja
  • zh
  • ko
  • ar
  • bn
  • bg
  • cs
  • da
  • nl
  • fi
  • ka
  • el
  • gu
  • he
  • hi
  • hu
  • id
  • kn
  • ms
  • ml
  • mr
  • nb
  • or
  • pl
  • pa
  • ro
  • ru
  • sk
  • sv
  • tl
  • ta
  • te
  • th
  • tr
  • uk
  • ur
  • vi

Platforms

System DownloadTested
Linux x86-64 19.0 MBDubbed with the real model from the release archive on AlmaLinux 8, and in CI
macOS (Apple silicon) 24.0 MBDubbed with the real model in CI (Apple Silicon)
Windows x86-64 18.5 MBDubbed with the real model in CI

Linux x86-64, 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 and dubs lines in six languages, a short line and a cloned 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: 1,115 MB, and 45 MB more for the speaker's own voice.
  • unsupported_languageThe project's language is not one of the 43, 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: open the video, or set a Fallback reference clip. The message says which.
  • out_of_memoryThe computer ran out of memory. Close other programs: ZONOS3-Mini needs about 2 GB free.
  • internalSomething failed inside the add-on; the message has the details. On an older processor without AVX2, the message says so.

License

Add-on
Apache-2.0
Engine
MIT
Zyphra's ZONOS3-Mini runtime, with ggml
Model
Apache-2.0
ZONOS3-Mini weights, by Zyphra
Voices
Apache-2.0
Zyphra's default voices
Text reading
MIT, Apache-2.0, Unicode-3.0
zyphra-textnorm, with grammars ported from NVIDIA's NeMo-text-processing and data from Unicode CLDR

The add-on is Apache-2.0. ZONOS3-Mini's weights and its default voices are Apache-2.0, and Zyphra's runtime is MIT, so commercial work is allowed. The text reading, zyphra-textnorm, is MIT, with grammars ported from NVIDIA's NeMo-text-processing (Apache-2.0) and data from Unicode CLDR.

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

  • 26.11.1

    First release. Offline text-to-speech with Zyphra's ZONOS3-Mini, on Zyphra's own ggml runtime, with no PyTorch.

    • 43 languages, at 44.1 kHz.
    • 27 stock voices (Zyphra's default voices). Every voice speaks every language, but with its own accent, which can be strong in another language; clone the speaker for those.
    • The speaker's own voice, cloned from 5 to 20 seconds of the video.
    • Numbers, dates, money and units are read out in each language's grammar (zyphra-textnorm).
    • Each line is checked as it is spoken (LACI, Zyphra's alignment check) and cut at the first silence once the text is read, so short lines don't run on with extra words. The cut is this add-on's own rule.
    • Runs on the CPU, about twice as fast as real time. x86-64 processors need AVX2.
    • Downloads: the model (1,115 MB) on the first line, the speaker encoder (45 MB) on the first cloned voice, pinned and checked by SHA-256.

    Linux builds run on glibc 2.28 and newer.

Other Voices add-ons