Download

Paradee (offline, lightweight English)

VoicesOfflineCPU onlyCommercial use OK

Dubs your subtitles in natural English with Paradee by Sahil Mahendrakar, an 8-million-parameter model distilled from Kokoro. Four voices, two American and two British, from one 8.5 MB model. It runs on your own computer's processor, many times faster than real time, with no graphics card and no account.

Install it from Subtitld

Settings Add-ons Paradee (offline, lightweight English) Install

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

Source on GitHub
Languages
2
Models
1
Voices
4
Add-on download
75–93 MB
RAM, at least
512 MB

What it does

Kokoro's voices, a tenth of the size

Paradee learned four of Kokoro's voices with a tenth of its parameters, and sounds close to it. Natural English at 24 kHz.

American and British

Heart and Puck speak American English, Isabella and Fable British English, each with its own pronunciation: "tomato", "schedule" and "can't" come out as each accent says them.

Fast on any computer

On a single processor core, a line comes out about 17 to 20 times faster than real time. About 300 MB of memory, and no graphics card.

Offline after the first line

The 8.5 MB model downloads the first time a line is spoken. After that, nothing leaves your computer. The models and the add-on allow commercial work.

Voices

In the Speakers tab, choose Paradee 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.

4 voices in 2 languages.

  • en-gb 2
  • en-us 2

Heart is the closest to Kokoro; Isabella is the furthest from it of the four. All four voices come from the same model, so switching voices downloads nothing. Paradee does not clone voices. To dub in the speaker's own voice, choose a voice-cloning engine such as Sopro or Chatterbox.

The model adds a little breath and hiss to its voices at random, as Kokoro does. Dubbing a line again can sound very slightly different there; the words, the voice and the timing stay the same.

Models

The model is not inside the add-on. It downloads into Subtitld's shared model cache the first time a line is spoken, 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
Paradee 1.2 (7.9M, int8 ONNX, four voices) Default paradee-1.2
8 MB
All four voices, about 17–20× real time on one processor core

The add-on runs the ONNX file that Paradee's author publishes, without PyTorch, and turns text into sounds exactly as the author's code does, with Kokoro's own phonemizer. We checked it against the author's code: the same sounds for every test sentence in both accents, and the same audio for all four voices.

Options

In Settings › Add-ons › Paradee (offline, lightweight English) › Configure.

CPU threads
Default: 0

0 uses one thread, which is already many times faster than real time. More threads rarely help with a model this small.

Speaking pace
Default: 1

How fast Paradee itself speaks: 1.0 is its natural pace, 1.2 is 20% faster. The model speaks faster or slower; the audio is not stretched.

Languages

English only. Any English project (American, British, Australian…) is accepted, and each voice reads it in its own accent. A project in another language gets a message that says so: pick another dubbing engine for it.

Numbers, prices, dates, times and common abbreviations are read out as words, and words such as "read" or "record" are said as their sentence needs. Names and words from other languages that the phonemizer doesn't know are said by espeak-ng, and may be mispronounced.

  • en-us
  • en-gb

Platforms

System DownloadTested
Linux x86-64 93.5 MBDubbed with all four voices on a real machine, in clean Debian, Ubuntu, Fedora and Arch systems, and in CI
macOS (Apple silicon) 75.2 MBDubbed with all four voices in CI (Apple Silicon)
Windows x86-64 84.5 MBDubbed with all four voices in CI

Linux x86-64 (any distribution from 2018 on), 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 with every voice.

Troubleshooting

  • network_unavailableThe model couldn't be downloaded, or a downloaded file failed its check and was discarded. Check your connection and try again: the download resumes.
  • disk_fullNot enough disk space for the model: 8.5 MB.
  • unsupported_languageThe project's language is not English, 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. Write the subtitle and dub it again.
  • out_of_memoryThe computer ran out of memory. Close other programs: Paradee needs about 400 MB.
  • internalSomething failed inside the add-on; the message has the details.

License

Add-on
GPL-3.0-or-later
Because it includes espeak-ng
Engine
Apache-2.0, MIT, GPL-3.0-or-later
misaki, Kokoro's own phonemizer, with spaCy turns the text into sounds; espeak-ng says the words they don't know; ONNX Runtime, by Microsoft, runs the model
Model
Apache-2.0
Paradee 1.2 weights, by Sahil Mahendrakar, distilled from Kokoro-82M by hexgrad

Fine for commercial work. The Paradee model and the author's code are Apache-2.0, like Kokoro. The add-on is GPL-3.0-or-later because it includes espeak-ng; its full source, espeak-ng's included, is on each release. The license covers the add-on, not what you make with it: the dubs you make are yours.

Versions

  • 26.11.1

    Offline English text-to-speech for Subtitld with Paradee 1.2: natural English speech at 24 kHz from an 8M-parameter model distilled from Kokoro-82M, many times faster than real time on the CPU. Four voices from one 8.5 MB model: Heart and Puck (American English), Isabella and Fable (British English).

    The archive holds the add-on, ONNX Runtime, misaki (Kokoro's phonemizer) with spaCy's small English pipeline, and espeak-ng with its English data. The model downloads the first time a line is spoken, at a pinned revision, checked by sha256.

    The add-on is GPL-3.0-or-later, because it bundles espeak-ng (GPL-3.0-or-later); the Paradee model and misaki are Apache-2.0, spaCy and its English pipeline MIT. Each archive carries THIRD_PARTY_LICENSES/ with the licence of everything frozen into it. The complete corresponding source is paradee-<version>-source.tar.gz in this release. The Linux archive also carries libquadmath (LGPL-2.1, with numpy); its source is paradee-<version>-libquadmath-source.tar.gz.

    What's Changed

    • Paradee add-on: lightweight offline English text-to-speech by @jonata in https://github.com/Subtitld/addon-paradee/pull/1
    • Parity test: allow other CPUs' int8 rounding by @jonata in https://github.com/Subtitld/addon-paradee/pull/2

    New Contributors

    • @jonata made their first contribution in https://github.com/Subtitld/addon-paradee/pull/1

Other Voices add-ons

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