tiny-voca offline, lightweight English

VoicesOfflineCPU onlyCommercial use OK

Dubs your subtitles in American English with tiny-voca, a 4.5-million-parameter model by Muhammed Nazeem, in a woman's or a man's voice. It runs on your own computer's processor, dozens of times faster than real time, and the model comes inside the add-on: nothing to download.

Install it from Subtitld

Settings Add-ons tiny-voca (offline, lightweight English) Install

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

Source on GitHub
Languages
1
Models
1
Voices
2
Add-on download
33–43 MB
RAM, at least
256 MB

Listen

What it does

Tiny, and clear

A 4.5-million-parameter model, one of the smallest that still sounds natural: clear American English at 24 kHz.

Nothing to download

The model, 26 MB, comes inside the add-on. It works the moment it is installed, and nothing leaves your computer.

Fast on any computer

A line comes out in a few hundredths of a second on a laptop processor, dozens of times faster than real time, with no graphics card.

Free for commercial work

The add-on and the model are Apache-2.0, and llama.cpp, which runs it, is MIT.

Voices

In the Speakers tab, choose tiny-voca 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.

2 voices in 1 language.

  • en 2

There are two voices, a woman and a man, both American English. tiny-voca's author made them from written descriptions, and trained the model on synthetic speech: no real person's voice was recorded. tiny-voca does not clone voices. To dub in the speaker's own voice, choose a voice-cloning engine such as Sopro or VoxMel.

Models

The model is inside the add-on, so there is nothing to download. It is the GGUF that ggml-org converted from tiny-voca, at a pinned revision, checked when the add-on was built.

Model SizeGood for
tiny-voca (4.5M) Default tiny-voca
25 MB
Both voices; part of the add-on

The model runs in llama.cpp, the same engine many people use for language models. llama.cpp also turns the text into sounds, with a port of tiny-voca's own phonemizer that comes inside the model.

Options

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

Speed
Default: 1

How fast tiny-voca itself speaks: 1.0 is its natural pace, 1.2 is 20% faster. A speaker's Rate in Subtitld changes it further. The model speaks faster or slower; the audio is not stretched.

CPU threads
Default: 0

0 uses up to four threads, which is already far faster than real time. More rarely help with a model this small.

Languages

English only. Any English project (American, British, Australian…) is accepted, read in an American accent. A project in another language gets a message that says so: pick another dubbing engine for it.

Numbers, prices, dates, times, units and common abbreviations are read out as words, and words such as "read" or "record" are said as their sentence needs.

  • en

Platforms

System DownloadTested
Linux x86-64 42.6 MBDubbed with both voices on a real machine and in clean AlmaLinux 8 and Ubuntu 24.04 systems; in CI, the installed add-on speaks with both voices
macOS (Apple silicon) 40.1 MBIn CI, the installed add-on speaks with both voices (Apple Silicon)
Windows x86-64 33.1 MBIn CI, the installed add-on speaks with both voices

Linux x86-64 (any distribution from 2018 on, glibc 2.28 or newer), macOS 13.3 or newer 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 and dubs with both voices.

Troubleshooting

  • unsupported_languageThe project's language is not English. 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.
  • internalSomething failed inside the add-on; the message has the details. Dubbing the line again restarts the speech engine.

License

Add-on
Apache-2.0
Engine
MIT
llama.cpp, which runs the model and turns the text into sounds
Model
Apache-2.0
tiny-voca weights, by Muhammed Nazeem; pronunciation data from dizione, the CMU Pronouncing Dictionary (BSD-2-Clause) and Google's Wikipedia Homograph Data

Fine for commercial work. The add-on and the tiny-voca model are Apache-2.0, and llama.cpp is MIT. The license covers the add-on, not what you make with it: the dubs you make are yours.

Versions

  • 26.11.1

    The first release: lightweight English text-to-speech for Subtitld with tiny-voca, a 4.5-million-parameter model by Muhammed Nazeem.

    • Two American English voices, a woman and a man, at 24 kHz. No voice cloning.
    • Dozens of times faster than real time on a CPU: a line takes a few hundredths of a second.
    • Numbers, prices, dates, times and units are read out as words.
    • The model (26 MB) is part of the add-on: nothing to download.
    • A Speed setting, which a speaker's Rate in Subtitld scales further.
    • English only: a line in another project language gets a message saying so.
    • Runs in llama.cpp's server, which the add-on starts on 127.0.0.1 and stops when it exits. The Linux archive runs on glibc 2.28 and newer.

    The add-on and the tiny-voca weights are Apache-2.0; llama.cpp and ggml are MIT. Each archive carries THIRD_PARTY_LICENSES/, including the licences of the pronunciation data in the model (CMU Pronouncing Dictionary, Wikipedia Homograph Data).

Other Voices add-ons