sanoTTS (ultra-light offline)

Offline neural text-to-speech at 294k-2.3M parameters — pure numpy, no torch, no GPU. 28 voices across 14 languages, each 0.3-9 MB, downloaded on first use and fully offline after that.

Voices

28 voices in 14 languages.

ar-jo 1
cs-cz 2
de-de 2
en-us 7
es-es 2
fr-fr 1
id-id 1
it-it 2
pt-br 2
ro-ro 2
ru-ru 2
tr-tr 2
vi-vn 1
zh-cn 1

Options

In Settings › Add-ons › sanoTTS (ultra-light offline) › Configure.

Speaking rate

Default: 1.0

Applied at synthesis time by stretching phoneme durations, which sounds better than resampling afterwards.

CPU threads

Default: 2

Threads numpy/OpenBLAS may use. The default of 2 keeps synthesis from saturating every core while you are editing.

Download voices from

Default: empty

Where voice packs are fetched from on first use. Pin one if the other is blocked on your network.

Local voice directory

Default: empty

Optional. Point at an already-extracted voice pack to synthesize with no network access at all.

Languages

ar-jo
cs-cz
de-de
en-us
es-es
fr-fr
id-id
it-it
pt-br
ro-ro
ru-ru
tr-tr
vi-vn
zh-cn

Platforms

System	Download
Linux x86-64	63.2 MB
macOS (Apple silicon)	37.6 MB

License

Add-on

GPL-3.0-or-later

GPL-3.0-or-later (add-on + bundled espeak-ng phonemizer); sanoTTS runtime is MIT; voice weights download on first use and are GPL-3.0

Versions

0.1.0

Source for the GPL components

This add-on is distributed under the GNU General Public License v3.0 or later. Complete Corresponding Source for the GPL-licensed parts (espeak-ng, phonemizer-fork) is attached to this release as *-corresponding-source.tar.gz, at no charge, per GPL-3.0 section 6(d).

Each binary archive also contains a THIRD_PARTY_LICENSES/ directory with the full licence texts and the exact upstream revisions used.