Dubs your subtitles on your own computer with Kokoro 82M, an Apache-2.0 speech model. Its 54 preset voices in nine languages all share one 82 MB model, downloaded the first time you use it. It can't clone a voice.

What it does

A voice per speaker

Choose Kokoro as the Dubbing engine and pick a voice for each speaker. Generate all speeches dubs that speaker's lines: all of them, a range, or the selection.

The voice sets the language

Each voice belongs to one language and accent, and Kokoro pronounces the text the way that language does. Pick a voice in the language of your subtitles.

Long lines in one clip

Kokoro speaks a long subtitle in pieces; the add-on joins them into one clip per subtitle.

Fits the subtitle

With Fit dub to subtitle duration on, each new clip is sped up or slowed down to last as long as its subtitle. The original clip is kept too.

Kokoro can't copy a voice from your video. For that, use Coqui XTTS, Qwen3-TTS or F5-TTS.

Voices

The Voice list shows all 54, each with its name, accent and gender, as in Bella — American female. Heart, first in the list, is marked as the default. American English has the most choice, with 20 voices; French has one.

54 voices in 9 languages.

en-gb 8
en-us 20
es 3
fr 1
hi 4
it 2
ja 5
pt-br 3
zh-cn 8

Model

Every voice uses the same model, Kokoro 82M. It isn't part of the add-on's download: Kokoro fetches it from Hugging Face (hexgrad/Kokoro-82M) the first time you dub, so that first line takes longer. The first line in each language also waits while Kokoro sets that language up; the lines after it don't.

Model	Download
kokoro-82m	82 MB

The model is kept in the Hugging Face cache, ~/.cache/huggingface/hub unless you've set HF_HOME. That's outside the add-on, so updating or uninstalling the add-on keeps it.

Options

In Settings › Add-ons › Kokoro 82M (offline) › Configure.

Inference device

Default: cpu

This version of the add-on doesn't pass this choice on to Kokoro, so changing it has no effect.

Speech speed multiplier (1.0 = normal)

Default: 1

Subtitld sends its own speaking rate with every line (normally 0) and the add-on uses that instead, so this setting has no effect in this version.

Languages

English comes in American (en-us) and British (en-gb) voices; every other language has one accent.

en-us
en-gb
es
fr
hi
it
ja
pt-br
zh-cn

Platforms

System	Download	Tested
Linux x86-64	384.9 MB	Start-up check in CI
macOS (universal)	275.6 MB	Start-up check in CI
Windows x86-64	255.8 MB	Start-up check in CI

Before every release, CI checks that each build starts and answers Subtitld. It doesn't synthesise speech there.

Troubleshooting

When a line fails, it's left without a dub. These are the reasons the add-on gives.

internalKokoro couldn't load, or couldn't speak the line. The first line in a language loads Kokoro, and the very first one downloads the model, so check your connection and try again.
unsupported_voiceThe voice saved for this line isn't a Kokoro voice. Pick one from the Voice list again.
bad_paramsThe subtitle has no text to speak.

License

Add-on

Apache-2.0

Engine

Apache-2.0

kokoro

Model

Apache-2.0

Kokoro 82M weights

Fine for commercial work: the add-on, the Kokoro engine and the model weights are all Apache-2.0.

Kokoro (offline)

What it does

Voices

Model

Options

Languages

Platforms

Troubleshooting

License

Versions

Other Voices add-ons